Paper deep dive
Unaligning Everything: Or Aligning Any Text to Any Image in Multimodal Models
Shaeke Salman, Md Montasir Bin Shams, Xiuwen Liu
Models: AltCLIP, BLIP-2, CLIP, CLIPSeg, ImageBind
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 94%
Last extracted: 3/12/2026, 5:33:06 PM
Summary
The paper presents a gradient-based adversarial attack method that exploits the shared embedding space of multimodal models (like ImageBind and CLIP). By minimally modifying input images, the authors demonstrate that they can align these images to arbitrary text embeddings, achieving a 100% success rate. This reveals a fundamental vulnerability where semantically unrelated images can be forced to match identical text embeddings, and visually indistinguishable images can be mapped to different text embeddings, regardless of the underlying classifier.
Entities (5)
Relation Signals (3)
Gradient-based optimization → attacks → Multimodal Model
confidence 95% · we show that we can align the embeddings of distinguishable texts to any image through unnoticeable adversarial attacks in joint image-text models
CLIP → uses → Shared Embedding Space
confidence 95% · CLIP (Radford et al. 2021) aligns vision and text representations.
ImageBind → exploits → Shared Embedding Space
confidence 90% · The shared embedding space could lead to new vulnerabilities if different modalities can be misaligned.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Utilizing a shared embedding space, emerging multimodal models exhibit unprecedented zero-shot capabilities. However, the shared embedding space could lead to new vulnerabilities if different modalities can be misaligned. In this paper, we extend and utilize a recently developed effective gradient-based procedure that allows us to match the embedding of a given text by minimally modifying an image. Using the procedure, we show that we can align the embeddings of distinguishable texts to any image through unnoticeable adversarial attacks in joint image-text models, revealing that semantically unrelated images can have embeddings of identical texts and at the same time visually indistinguishable images can be matched to the embeddings of very different texts. Our technique achieves 100\% success rate when it is applied to text datasets and images from multiple sources. Without overcoming the vulnerability, multimodal models cannot robustly align inputs from different modalities in a semantically meaningful way. \textbf{Warning: the text data used in this paper are toxic in nature and may be offensive to some readers.}
Tags
Links
Trouble viewing inline? Open PDF directly →
Full Text
112,380 characters extracted from source content.
Expand or collapse full text
Unaligning Everything: Or Aligning Any Text to Any Image in Multimodal Models Shaeke Salman 1 , Md Montasir Bin Shams 1 , Xiuwen Liu 1 1 Department of Computer Science, Florida State University, FL 32306, USA salman, liux@cs.fsu.edu, mshams@fsu.edu Abstract Utilizing a shared embedding space, emerging multimodal models exhibit unprecedented zero-shot capabilities. How- ever, the shared embedding space could lead to new vulnera- bilities if different modalities can be misaligned. In this paper, we extend and utilize a recently developed effective gradient- based procedure that allows us to match the embedding of a given text by minimally modifying an image. Using the pro- cedure, we show that we can align the embeddings of distin- guishable texts to any image through unnoticeable adversar- ial attacks in joint image-text models, revealing that semanti- cally unrelated images can have embeddings of identical texts and at the same time visually indistinguishable images can be matched to the embeddings of very different texts. Our technique achieves 100% success rate when it is applied to text datasets and images from multiple sources. Without over- coming the vulnerability, multimodal models cannot robustly align inputs from different modalities in a semantically mean- ingful way.Warning: the text data used in this paper are toxic in nature and may be offensive to some readers. Introduction Built on large pre-trained foundation models (Bommasani et al. 2022), applications have exhibited unprecedented ca- pabilities for a wide range of tasks, setting new state-of- the-art on benchmark datasets, acing standard exams, and passing professional exams (OpenAI 2023; Brandes et al. 2022; Kung et al. 2023; Choi et al. 2023). Such models, however, are not well understood due to their complexity, even though the need for understanding and the risks of lack- ing is widely recognized and acknowledged (Bommasani et al. 2022). For example, transformers have become a hall- mark component in models for many applications and have led to significant improvements in performance on bench- mark datasets (Vaswani et al. 2017; Dosovitskiy et al. 2021; Devlin et al. 2019). By transforming inputs from different modalities (such as texts and images) to a common embed- ding space, emerging multimodal models provide new capa- bilities and new applications are being developed by exploit- ing the shared space (Radford et al. 2021). At the same time, it is well known that neural networks ex- hibit an intriguing property in that they are subject to adver- sarial attacks: some small changes to an input could result in substantial changes in model responses and outputs (Good- fellow, Shlens, and Szegedy 2015; Szegedy et al. 2014; Chakraborty et al. 2021). While studies have shown adver- sarial examples exist to break even aligned models (Zou et al. 2023), it is not clear whether the shared space could be exploited to establish arbitrary associations between im- ages and texts, or between two different modalities, there- fore breaking the alignments that many of the models rely on in order to function properly. In this paper, using a gradient-descent-based optimization procedure as detailed in our prior work (Salman, Shams, and Liu 2024), we show that perturbing an input image to a deployed model in unnoticeable ways can alter the result- ing representation to match any chosen text and, therefore, reveal an inherent vulnerability of joint vision-text models. Since such multimodal models are being deployed, the iden- tified vulnerability should be considered for these models. Furthermore, we show that the resulting inputs can dramat- ically change classification results with no modifications to the classifiers. To highlight the main advantages of our framework, we present our results using multiple models, including the Im- ageBind (Girdhar et al. 2023). Fig. 1 shows several images along with their representations and the classification re- sults. The three visually indistinguishable images in the top row of Fig. 1 (see Fig. 9 in appendix for pixel differences) have very different representations, as shown by their low- dimensional projections; the images in the bottom row also have very different representations. On the other hand, the pairs in each of the three columns in Fig. 1 have very similar representations even though they are semantically very dif- ferent. When we pass these images to the unmodified mul- timodal ImageBind model, the images with similar embed- dings are classified into the same class, regardless of their semantic similarity, as shown in Fig. 1 (d) and (h). These and additional results shown in the Experiments sec- tion, along with the fact we have obtained the same find- ings on all the image-text pairs we have used, demonstrate convincingly that there are visually indistinguishable inputs corresponding to the embeddings of very different texts, and yet there are very different images corresponding to the em- beddings of identical texts. By analyzing the equivalence classes (Salman, Shams, and Liu 2024) of the embeddings arXiv:2407.01157v1 [cs.CV] 1 Jul 2024 strawberry -> "loser" 012345 principal component 6 3 0 3 6 9 projected value (a) embedding projectionstrawberry -> "ignorant" 012345 principal component 6 3 0 3 6 9 projected value (b) embedding projectionstrawberry -> "ugly" 012345 principal component 6 3 0 3 6 9 projected value (c) embedding projection (d) Image (a) (b) (c) "loser""ignorant""ugly" 1.0 7.3684e-23 5.25e-20 3.1527e-21 1.0 2.3794e-21 1.5667e-18 2.5333e-20 1.0 vision x text cauliflower -> "loser" 012345 principal component 6 3 0 3 6 9 projected value (e) embedding projectioncauliflower -> "ignorant" 012345 principal component 6 3 0 3 6 9 projected value (f) embedding projectioncauliflower -> "ugly" 012345 principal component 6 3 0 3 6 9 projected value (g) embedding projection (h) Image (e) (f) (g) "loser""ignorant""ugly" 1.0 1.6182e-22 6.2921e-20 5.4233e-21 1.0 5.2222e-21 3.4878e-18 3.3504e-20 1.0 vision x text Figure 1: Typical examples from ImageNet obtained using the proposed framework. The visually indistinguishable images have different representations from each other as shown in their low-dimensional projections. Note that the arrow in the title (original→target) signifies a derived image from the original one by aligning the embedding of the original image with the target text embedding using our method. The projections of embedding-aligned images closely resemble the projections of the aligned text. The matrix shows the classification outcomes from the multimodal ImageBind pretrained model used directly with no modifications; each row corresponds to one image. of vision-text models, the vulnerability we show is due to the representations used by such models and do not depend on application-specific classifiers. Our main contributions are as follows: • Using our efficient computational procedure to match specified representations, we clearly show an inherent vulnerability of joint vision-text models, where arbitrary associations between the images and texts can be estab- lished, regardless of the semantics of the images and texts. More specifically, we show that visually indistin- guishable images can have very different representations, yet unrelated images semantically can correspond to sim- ilar text embeddings. • We show that we can map all the texts in differ- ent datasets to visually indistinguishable images with a 100% success rate. Related Work The transformer architecture (Vaswani et al. 2017) revolu- tionized NLP by effectively capturing long-range dependen- cies, resulting in powerful pre-trained models like BERT (Devlin et al. 2019) and GPT (Brown et al. 2020) that excel in various tasks. This advancement extends to computer vi- sion with the Vision Transformer (ViT) (Dosovitskiy et al. 2021), showcasing the transformative impact of attention mechanisms across domains. The recent prompting-based models and multimodal mod- els have further accelerated the trend. The joint multimodal models have demonstrated significant benefits by employ- ing a shared embedding space across various modalities. For example, CLIP (Radford et al. 2021) aligns vision and text representations. The model is trained to predict the co- herence of image-text pairs, which enables it to understand complex relationships between the two modalities. Several recent works extend the technique of shared embedding space beyond text and vision by employing a unified em- beddings space in various modalities are: GPT-4 (OpenAI 2023), MiniGPT-4 (Zhu et al. 2023), Flamingo (Alayrac et al. 2022), Bard (Pichai 2023), LLaVA (Liu et al. 2023) and, ImageBind (Girdhar et al. 2023). Another line of research aims to comprehend models by probing them to unveil new properties. A well-studied prob- lem is adversarial attacks, where unnoticeable changes to the input can cause the models, primarily classifiers, to change their predictions. Most adversarial attacks are ap- plied to commonly used (deep) neural networks, includ- ing multiple-layer perceptrons and convolutional neural net- works (CNNs), demonstrating their vulnerability and sen- sitivities to such adversarial changes (Szegedy et al. 2014; Goodfellow, Shlens, and Szegedy 2015). Croce and Hein propose AutoAttack, an ensemble of parameter-free attacks that combines multiple methods to provide a robust assess- ment of a model’s vulnerability (Croce and Hein 2020). Re- cent studies have explored the vulnerability of multimodal models to adversarial attacks, which can potentially jail- break aligned large language models (LLMs) or Vision Lan- guage models (VLMs) (Carlini et al. 2023; Qi et al. 2023; Zou et al. 2023). Bhojanapalli et al. investigate the robust- ness of ViTs against attacks where the attacker has access to the model’s internal structure (Bhojanapalli et al. 2021; Shao et al. 2022). Notably, these methods revolve around gener- ating adversarial examples based on the classifier’s method- ology rather than focusing on the representation level. How- ever, our approach is different. Rather than crafting an adver- sarial example tailored to a particular classifier, our method is designed to generate examples that conform to a specified representation. A closely related study by Kazemi et al. examines the be- havior and vulnerabilities of the CLIP model (Kazemi et al. 2024). Their work is centered on inverting the CLIP model embeddings to understand the semantic information of these embeddings. However, our work focuses on demonstrating the vulnerabilities within the shared embedding spaces of multimodal models through adversarial attacks. We show through extensive experiments that visually indistinguish- able images can be mapped to arbitrary text, revealing an inherent vulnerability that is classifier agnostic. Preliminaries As this paper focuses on vision-language models that are based on transformers, here we first describe the transformers mathematically and then describe the vision- language models. Transformers can be described mathe- matically succinctly, consisting of a stack of transformer blocks. A transformer block is a parameterized func- tion classf θ :R n×d →R n×d . Ifx∈R n×d then f θ (x) =zwhereQ (h) (x i ) =W T h,q x i , K (h) (x i ) = W T h,k x i , V (h) (x i ) =W T h,v x i , W h,q ,W h,k ,W h,v ∈ R d×k . The key multi-head self-attention is a softmax func- tion applying row-wise on the inner products. 1 α (h) i,j =softmax j Q (h) (x i ),K (h) (x j ) √ k ! .(1) The outputs from the softmax are used as weights to com- pute new features, emphasizing the ones with higher weights given by u ′ i = H X h=1 W T c,h n X j=1 α i,j V (h) (x j ), W c,h ∈R k×d .(2) The new features then pass through a layer normalization, followed by a ReLU layer, and then another layer normaliza- tion. Typically transformer layers are stacked to form deep models. While transformer models are widely used for natural lan- guage processing tasks, recently, they are adapted to vision tasks by using image blocks on the basic units, and spa- tial relationships among the units are captured via the self- attention mechanism (Dosovitskiy et al. 2021). A vision- language model based on transformers incorporates a ded- icated transformer model for each input modality. The re- sulting representations from these modalities are mapped to a shared embedding space. For the ImageBind model, we denote the model for image xasf I (x)and for texttasf T (t). Fig. 2 shows that the image embeddings and text embeddings share the same vec- tor space. Given imagexandCtext labels,t 0 ,...,t C−1 , 1 Note that there are other ways to compute the attention weights. 7.55.02.50.02.55.07.510.0 first principal component 6 4 2 0 2 4 6 second principal component image_embeddings text_embeddings Figure 2: Low-dimensional projections of the embeddings of images and texts, showing texts and images share the same embedding space. We use all of the toxic comments (i.e., 992) from the 1,2,3-tokens toxic dataset and the same num- ber of strawberry and cauliflower images from ImageNet. the zero-shot classification uses softmax applied on the dot products of the image and text representations. Therefore, the probability classifiedxtot i is given by e f I (x) T f T (t i ) P C−1 j=0 e f I (x) T f T (t j ) ,(3) which is a typical implementation of the softmax function. Note that the probabilities reported in this paper’s figures are computed using a publicly available ImageBind model without any change. While the proposed method applies to all transformer-based models with continuous inputs, we fo- cus on multiple models, including the CLIP model (Radford et al. 2021), which jointly models images and text using the same shared embedding space as the ImageBind (Girdhar et al. 2023) model. Ideally, only images and texts that are semantically related should have similar embeddings. The image and text embeddings can help each other, resulting in robust zero-shot capability (Radford et al. 2021). On the other hand, vulnerabilities in associations between images and texts could be exploited, resulting in new weaknesses. Methodology Understanding the structures of the representation space is crucial for determining how the model generalizes. As intro- duced in our earlier work (Salman, Shams, and Liu 2024), we have proposed a framework to explore and analyze the embedding space of vision transformers, uncovering intrigu- ing equivalence structures and their implications for model generalization and robustness. Generally, we model the rep- resentation given by a (deep) neural network (including a transformer) as a functionf:R m →R n . A fundamen- tal question is to have a computationally efficient and ef- fective way to explore the embeddings of inputs by finding the inputs whose representation will match the one given by f(x tg ), wherex tg is an input whose embedding we like to match. Informally, given an image of a strawberry in Fig. 1 as an example, all the images that share its representation given by a model will be treated as a strawberry. Embedding Alignment Procedure As described in our previous work (Salman, Shams, and Liu 2024), the proposed approach for embedding alignment fo- cuses on aligning the representation of an input with that of a target input. The core of the method is an iterative gradient optimization procedure, similar to how most of the neural networks are trained, except the gradient is calculated with respect to the input variables. Since we need to match two vectors, we define the loss for finding an input matching a given representation as L(x) =L(x 0 + ∆x) = 1 2 ∥f I (x 0 + ∆x)−f T (t tg )∥ 2 ,(4) wherex 0 is an initial image andf T (t tg )specifies the target embedding for a specified text sequencet tg . Approximately, the gradient is given by ∂L ∂x ≈ ∂f ∂x x=x 0 T (f I (x 0 + ∆x)−f T (t tg )).(5) In each step, the algorithm first calculates the loss as defined by the loss function, between the image embedding with the target embedding as the one given by the specified text. Then using PyTorch, it computes the gradient by doing backward computation. After the gradient is calculated, we update the pixel values by doing gradient descent. Eq. (5) shows how the gradient of the mean square loss func- tion is related to the Jacobian of the representation function atx=x 0 . In other words, the gradient is related to the dif- ferences between the current and target text representations. When the differences are large, the gradient should be sig- nificant as well. Consistent with this analysis, on all the ex- amples we have experimented with, we are able to minimize the loss, and details are provided in the Experimental Results section and appendix. While local optimal solutions could be obtained by solving a quadratic programming problem or linear programming problem, depending on the norm to be used when minimiz- ing∆x, the gradient function works effectively for all the cases we have tested due to the Jacobian of the transformer. One of the practical issues using the gradient descent-based procedure is how to determine the learning rate. In the case of the transformers, the model can be approximated by a lin- ear model when it moves within one activation region; note that it is approximate due to the nonlinearity of the softmax. The algorithm is able to find the matching representations for a wide range of learning rates; see the Experimental Re- sults section for more details. Experiments In this section, we first outline the specifics of our exper- imental settings and implementation details. Our designed framework is systematically applied across various datasets and multiple multimodal models; in the subsequent subsec- tions, we present both the experimental outcomes and quan- titative results. Our findings showcase the capability to align any distinguishable text with an image through impercepti- ble adversarial attacks within a joint image-text model. More importantly, we show that our framework exhibits versatil- ity, being agnostic to both the model architecture and dataset characteristics. Datasets and Settings Datasets.We conduct extensive experiments to evaluate our proposed framework on widely recognized publicly avail- able vision datasets, namely ImageNet (Deng et al. 2009) and MS-COCO (Lin et al. 2015). For the text aspect, we adopt a methodology inspired by the work of Jones et al. (2023) and Borkan et al. (2019), obtaining a dataset com- prising 68, 332, and 592 toxic comments with 1, 2, and 3 to- kens respectively 2 . Additionally, our framework undergoes evaluation using the Jigsaw toxic dataset available at Kag- gle 3 (van Aken et al. 2018). Implementation Details.To demonstrate the feasibility of the proposed method on large multimodal models, we have used the pretrained model publicly available by ImageBind 4 , which in turn uses a CLIP model. 5 More specifically, Image- Bind utilizes the pre-trained vision (ViT-H 630M params) and text encoders (302M params) from the OpenCLIP (Il- harco et al. 2021; Girdhar et al. 2023). The input size is 224×224×3, and the dimension of the embedding is 1024. We perform all our experiments on a lab workstation featur- ing two NVIDIA A5000 GPUs. We will provide source code for all our experiments in GitHub 6 . Additional Models.To demonstrate the broader applica- bility of the models, we thoroughly evaluate with several other multimodal models, including CLIPSeg (L ̈ uddecke and Ecker 2022), AltCLIP (Chen et al. 2022), BLIP-2 (Li et al. 2023), etc. An example with CLIPSeg is presented in the following subsection. The size of the input is determined by the models. For all the multimodal models we have used, a preprocessing step is used to resize the input image to 224 × 224 × 3 for subsequent processing. Therefore, the method works equally well regardless of the resolution of the origi- nal input images. Embedding Projections.To obtain the low-dimensional projections of the images shown in Fig. 1 and other simi- lar figures, the largest principal components are computed from a subset of images from the ImageNet dataset. Then, we project an embedding to be displayed along the six prin- cipal components with the largest eigenvalues. Note that the projections are used to illustrate the differences between em- beddings, and details of the principal components would not impact the results significantly in that similar embeddings will have similar projections, and different embeddings will have different projections. 2 https://github.com/ejones313/auditing-llms/tree/main/data 3 https://w.kaggle.com/c/jigsaw-toxic-comment- classification-challenge 4 https://github.com/facebookresearch/ImageBind 5 https://github.com/mlfoundations/open clip 6 https://github.com/programminglove08/UnalignMM racing car -> "utterly pathetic post"racing car -> "what naive stupidity"racing car -> "a dumb move"racing car -> "only fools disagree"racing car -> "you are disgusting" 1.02.3198e-233.6043e-202.5487e-221.5429e-22 6.9174e-231.02.2489e-204.5092e-232.4291e-24 8.2754e-203.0503e-191.02.1984e-226.4248e-21 1.103e-222.6186e-266.5155e-251.02.1915e-20 2.0849e-231.2924e-244.6186e-223.5099e-181.0 vision x text cat -> "you are disgusting"dog -> "you are disgusting"flowerpot -> "you are disgusting"panda -> "you are disgusting"zebra -> "you are disgusting" 4.2102e-234.5288e-252.183e-229.8592e-191.0 2.089e-224.1558e-247.7363e-221.4802e-191.0 1.0965e-215.8426e-239.9812e-212.1077e-181.0 1.3588e-232.4474e-249.803e-222.6144e-201.0 5.2986e-238.2344e-254.1938e-234.5226e-191.0 vision x text Figure 3: (top) More examples involving ImageNet and 1,2,3-tokens toxic dataset, where visually indistinguishable images have very different representations via embedding alignment with the corresponding texts and therefore very different classification outcomes (as shown in the classification probabilities; each row in the matrix corresponds to one image (from left to right)). (bottom) Visually very different images have very similar embeddings, aligned and classified to a particular text. 02000400060008000 steps 0.1 0.2 0.3 0.4 0.5 0.6 0.7 loss 02000400060008000 steps 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 cosine similarity cosine_similarity avg_diff 0.000 0.005 0.010 0.015 0.020 0.025 0.030 0.035 0.040 average pixel differance Figure 4: The evolution of loss while matching a target em- bedding. (left) the loss w.r.t. steps. (right) the cosine similar- ity between the embeddings of the new input and the target w.r.t. the steps, along with the average pixel value difference between the new input and the original image. Experimental Results To demonstrate the effectiveness of our method on deployed models such as the CLIP model, a key is to be able to match a given representation given by a phrase, a sentence, or any text sequence that can be encoded by the text transformer. We have tested the embedding matching procedure using many image and text pairs and Fig. 4 shows a typical exam- ple. The left of Fig. 4 shows the evolution of the loss when matching the embedding of an image to a specified target embedding. Similar to gradient descent, where the loss could become higher or lower, the high peaks indicate noise dur- ing the optimization process because the loss is non-linear. We use a small step size to make sure it converges. The right panel shows that cosine similarity increases steadily. We also show the average pixel value difference between the new in- put and the original image at each step; one can see the val- ues remain very small even though they increase as well. The algorithm is not sensitive to the learning rate and works effectively across a broad range of values, spanning from 0.001 to 0.09. For instance, with a learning rate of 0.001, 0.00.20.40.60.81.0 cosine similarity 0 10 20 30 40 density texts_1 texts_2 aligned image-texts Figure 5: (to be viewed in color) Cosine similarity distribu- tion. The red and green ones stand for the cosine similar- ity values corresponding to pairs of texts (i.e., embeddings) from the two toxic datasets considered. The blue one shows the distribution of cosine similarities of the embeddings of embedding-aligned image and text pair from the ImageNet and toxic dataset. As the cosine similarities of toxic data pairs do not overlap with other embeddings, potential map- ping opportunities exist. convergence is achieved in around 40,000 iterations, while 0.09 requires around 8,000 iterations. The visual differences in the resulting images are not noticeable. Eqn. 4 and 5 pro- vide an explanation, as the gradient for our loss is insensitive to the learning rate. Systematic evaluation.To further demonstrate the effec- tiveness of the gradient procedure to match embeddings, we have applied them to numerous images and texts from dif- ferent sources. Understanding the algebraic and geometric structures of the embedding space allows us to explore the strawberry -> "loser" 012345 principal component 6 3 0 3 6 9 projected value (a) embedding projectionstrawberry -> "ignorant" 012345 principal component 6 3 0 3 6 9 projected value (b) embedding projectionstrawberry -> "ugly" 012345 principal component 6 3 0 3 6 9 projected value (c) embedding projection (d) Image (a) (b) (c) "loser""ignorant""ugly" 1.0 9.5749e-07 1.0697e-05 8.3426e-07 1.0 9.4544e-06 1.4285e-06 1.0243e-06 0.99998 vision x text cauliflower -> "loser" 012345 principal component 6 3 0 3 6 9 projected value (e) embedding projectioncauliflower -> "ignorant" 012345 principal component 6 3 0 3 6 9 projected value (f) embedding projectioncauliflower -> "ugly" 012345 principal component 6 3 0 3 6 9 projected value (g) embedding projection (h) Image (e) (f) (g) "loser""ignorant""ugly" 1.0 7.285e-07 1.2405e-05 1.0391e-06 1.0 5.5667e-06 1.6406e-06 1.7561e-06 0.99998 vision x text Figure 6: Examples obtained using the proposed framework for different multimodal models, such as CLIPSeg. The results are given in the same format as depicted in Fig. 1. The example demonstrates that the method is model-agnostic. stop sign -> "speed limit 70"road intersection -> "speed limit 70"bridge -> "speed limit 70"building -> "speed limit 70"tunnel -> "speed limit 70" 4.6813e-073.5065e-108.8599e-111.8366e-121.4254e-111.0 3.1261e-082.664e-106.4921e-111.8939e-122.4207e-111.0 2.4771e-082.5447e-102.6075e-102.0127e-123.741e-111.0 4.7015e-082.7203e-108.2826e-117.5318e-123.0061e-111.0 5.1961e-081.2581e-104.5043e-111.0953e-125.1384e-111.0 vision x text Figure 7: Real-world scenarios where the proposed method is applied. The images are taken randomly from the web. All the matched images (for a stop sign, a road intersection, a bridge, a building, and a tunnel) are recognized as the sign “speed limit 70”. The examples demonstrate that our method works robustly for any data examples, therefore, the method is dataset-agnostic. space effectively. For example, we can find adversarial at- tacks to the embedding of any given image or text using the proposed gradient procedure. Fig. 1 shows two examples. To demonstrate the universal applicability of the procedure and the adversarial examples that exist almost everywhere, Fig. 3 shows more examples from different categories from the ImageNet dataset. See the appendix for additional ex- amples on ImageNet and MS-COCO datasets. The efficacy of the procedure is model-agnostic. To substantiate and con- firm this, Fig. 6 illustrates an example of a different multi- modal model using CLIPSeg, showcasing the consistent ap- plication and effectiveness of the approach irrespective of the specific model employed. In addition, Fig. 7 shows real- world scenarios, demonstrating the practical relevance of our findings. All these examples convincingly demonstrate our method is model and dataset-agnostic. Quantitative evaluation.Fig. 5 depicts the distribution of cosine similarities: the red and green curves represent co- sine similarity values corresponding to pairs of text embed- dings from the two toxic datasets under consideration re- spectively. The blue curve illustrates the distribution of co- sine similarities between embeddings of image and text pairs aligned through our embedding process from the ImageNet DataMatch Success RateMeanℓ 2 Distortion 1-Token100%0.98±0.09 2-Token100%0.83±0.15 3-Token100%0.47±0.11 Table 1: Success rates and meanℓ 2 distortions after we align the embeddings of given images (from ImageNet) to all the 1, 2, and 3-token toxic comments, respectively, in the toxic dataset. and toxic dataset. Essentially, the absence of overlap indi- cates that we can subtly modify an image corresponding to any selected text. In other words, with our approach, if we are provided with two or more texts, we can generate multi- ple visually indistinguishable images, one for each text, en- suring that a classifier will classify every image to the as- signed text, regardless of the semantics of the images. Due to this characteristic, Table 1 demonstrates a100%success rate (Carlini et al. 2023) in accurately matching the images with the toxic texts. To define the success rate, we first estab- lish criteria for a successful image alignment. After aligning an image with the embedding of a specific text, we utilize the imageBind model for classification. If the resulting clas- sification matches the given text, the alignment is considered successful; otherwise, it is not. The success rate on a partic- ular dataset is then calculated as the percentage of images that meet these criteria. As a concrete example, Fig. 14 il- lustrates the alignment of each of the 68 different text em- beddings with an image. Since all cases are successful, the success rate is 100%. In addition to metrics such as attack success rate and meanℓ 2 distortion shown in Table 1, we note the number of pixels changed above specified thresh- olds (>0.03 : 4500|>0.05 : 795|>0.08 : 90|>0.1 : 25|>0.2 : 1) and the meanℓ ∞ norm of the difference im- ages (0.09, 0.07, and 0.03 for 1, 2, and 3-token comments, respectively), as these additional evaluation metrics are com- monly used. Adversarial Modification Detection:We have observed that the embedding-matched images exhibit much higher sensitivity to Gaussian noise than the original ones. Lever- aging this insight, we have designed a detection algorithm introduced in our previous work (Salman et al. 2024). As demonstrated in that study, the detection algorithm performs reliably and consistently across a wide range of standard de- viations. The process is as follows: we add Gaussian noise of a specified standard deviation to a given image and then classify them. If the labels of the two images agree, the im- age is unmodified; otherwise, the image is modified. Discussion and Future Work It may be attempting to categorize our framework as an ad- versarial attack technique. Our primary focus is on analyzing the embedding space; we utilize the ImageBind solely as a classifier to validate our findings and is not used otherwise. While our embedding matching procedure can be used to generate effective adversarial examples, it is fundamentally different. Our technique is classifier agnostic and does not exploit features specific to classifiers. Consequently, our ex- amples with matched embeddings will appear to be the same to any classifier or downstream model that builds on embed- dings. On the other hand, traditional adversarial attacks are specific to classifiers and applications, focusing on altering their outputs by changing the input. Identifying adversarial attacks on multimodal models is very active (Qi et al. 2023; Schlarmann and Hein 2023; Evtimov et al. 2021). In general, all of them focus on how small changes in inputs can alter the final output (such as cap- tions or classification labels) across various models. In con- trast, our work identifies a new representation vulnerabil- ity. For instance, as shown in Fig. 1, three strawberry and cauliflower pairs can be made to be associated with three different texts, highlighting a more foundational vulnerabil- ity. The plausible root cause of such adversarial examples and also semantically different images with identical embed- dings is that transformers do not require the inputs to be aligned to have similar embeddings. By adding alignment- sensitive components to the embedding could mitigate the problem, which is being investigated further. Given that the models are susceptible to such adversarial at- tacks, a logical question is if there is an effective method to mitigate the attacks. One potential way to do so is to train a model further to reduce the vulnerabilities. For deep neu- ral networks, robust adversarial training has been used with success (Bai et al. 2021). It is unclear how much an adver- sarially trained model will affect the algorithm’s ability to match images with text, and this is currently being investi- gated. The results shown in this paper seem not to be consistent with the impressive results demonstrated by such models. Note that almost all existing results are measured on bench- mark datasets. Due to the high dimensionality of the embed- ding space and the input space, even the largest dataset will cover the spaces very sparsely. We believe that systematic evaluations such as ours are necessary if one likes to evalu- ate models to be able to predict their behaviors in the entire space rather than on samples. Conclusion In this paper, using a gradient descent-based procedure, we have revealed a new vulnerability in multimodal models, where semantically unrelated inputs can have similar rep- resentations, and, at the same time, semantically identical images can have very different representations. Therefore, aligning different inputs to shared embedding space in a se- mantically meaningful way may not be viable. As multiple models are being developed, one must consider the vulner- abilities in multimodal models for secure applications. As the proposed technique can associate any image with any chosen text, one must understand the implications of this in- herent vulnerability. References Alayrac, J.-B.; Donahue, J.; Luc, P.; Miech, A.; Barr, I.; Has- son, Y.; Lenc, K.; Mensch, A.; Millican, K.; Reynolds, M.; et al. 2022. Flamingo: a visual language model for few- shot learning. InAdvances in Neural Information Process- ing Systems, volume 35, 23716–23736. Bai, T.; Luo, J.; Zhao, J.; Wen, B.; and Wang, Q. 2021. Re- cent advances in adversarial training for adversarial robust- ness.arXiv preprint arXiv:2102.01356. Bhojanapalli, S.; Chakrabarti, A.; Glasner, D.; Li, D.; Un- terthiner, T.; and Veit, A. 2021. Understanding Robustness of Transformers for Image Classification. InIEEE/CVF In- ternational Conference on Computer Vision (ICCV). Bommasani, R.; Hudson, D. A.; Adeli, E.; Others; and et al. 2022. On the Opportunities and Risks of Foundation Mod- els.CoRR. Borkan, D.; Dixon, L.; Sorensen, J.; Thain, N.; and Vasser- man, L. 2019.Nuanced Metrics for Measuring Un- intended Bias with Real Data for Text Classification. arXiv:1903.04561. Brandes, N.; Ofer, D.; Peleg, Y.; Rappoport, N.; and Linial, M. 2022. ProteinBERT: a universal deep-learning model of protein sequence and function.Bioinformatics, 38(8): 2102– 2110. Brown, T. B.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; Agarwal, S.; Herbert-Voss, A.; Krueger, G.; Henighan, T.; Child, R.; Ramesh, A.; Ziegler, D. M.; Wu, J.; Winter, C.; Hesse, C.; Chen, M.; Sigler, E.; Litwin, M.; Gray, S.; Chess, B.; Clark, J.; Berner, C.; McCandlish, S.; Radford, A.; Sutskever, I.; and Amodei, D. 2020. Language Models are Few-Shot Learners. arXiv:2005.14165. Carlini, N.; Nasr, M.; Choquette-Choo, C. A.; Jagielski, M.; Gao, I.; Awadalla, A.; Koh, P. W.; Ippolito, D.; Lee, K.; Tramer, F.; and Schmidt, L. 2023. Are aligned neural net- works adversarially aligned?CoRR. Chakraborty, A.; Alam, M.; Dey, V.; Chattopadhyay, A.; and Mukhopadhyay, D. 2021. A survey on adversarial attacks and defences.CAAI Transactions on Intelligence Technol- ogy, 6(1): 25–45. Chen, Z.; Liu, G.; Zhang, B.-W.; Ye, F.; Yang, Q.; and Wu, L. 2022. AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities. arXiv:2211.06679. Choi, J. H.; Hickman, K. E.; Monahan, A.; and Schwarcz, D. B. 2023. ChatGPT Goes to Law School.Journal of Legal Education (Forthcoming). Croce, F.; and Hein, M. 2020. Reliable evaluation of adver- sarial robustness with an ensemble of diverse parameter-free attacks. arXiv:2003.01690. Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei- Fei, L. 2009. ImageNet: A large-scale hierarchical image database. In2009 IEEE Conference on Computer Vision and Pattern Recognition, 248–255. Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. InProceedings of NAACL-HLT 2019, 4171–4186. Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; Uszkoreit, J.; and Houlsby, N. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. InInternational Conference on Learn- ing Representations. Evtimov, I.; Howes, R.; Dolhansky, B.; Firooz, H.; and Ferrer, C. C. 2021.Adversarial Evaluation of Mul- timodal Models under Realistic Gray Box Assumption. arXiv:2011.12902. Girdhar, R.; El-Nouby, A.; Liu, Z.; Singh, M.; Alwala, K. V.; Joulin, A.; and Misra, I. 2023. ImageBind: One Embed- ding Space To Bind Them All. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 15180– 15190. Goodfellow, I. J.; Shlens, J.; and Szegedy, C. 2015. Explain- ing and Harnessing Adversarial Examples. InInternational Conference on Learning Representations. Hor ́ e, A.; and Ziou, D. 2010. Image Quality Metrics: PSNR vs. SSIM. In2010 20th International Conference on Pattern Recognition, 2366–2369. Ilharco, G.; Wortsman, M.; Wightman, R.; Gordon, C.; Car- lini, N.; Taori, R.; Dave, A.; Shankar, V.; Namkoong, H.; Miller, J.; Hajishirzi, H.; Farhadi, A.; and Schmidt, L. 2021. OpenCLIP. Jones, E.; Dragan, A.; Raghunathan, A.; and Steinhardt, J. 2023. Automatically Auditing Large Language Models via Discrete Optimization. arXiv:2303.04381. Kazemi, H.; Chegini, A.; Geiping, J.; Feizi, S.; and Gold- stein, T. 2024. What do we learn from inverting CLIP mod- els? arXiv:2403.02580. Kung, T. H.; Cheatham, M.; Medenilla, A.; Sillos, C.; De Leon, L.; Elepa ̃ no, C.; Madriaga, M.; Aggabao, R.; Diaz- Candido, G.; Maningo, J.; and Tseng, V. 2023. Performance of ChatGPT on USMLE: Potential for AI-assisted medi- cal education using large language models.PLOS Digital Health. Li, J.; Li, D.; Savarese, S.; and Hoi, S. 2023. BLIP-2: Boot- strapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. arXiv:2301.12597. Lin, T.-Y.; Maire, M.; Belongie, S.; Bourdev, L.; Girshick, R.; Hays, J.; Perona, P.; Ramanan, D.; Zitnick, C. L.; and Doll ́ ar, P. 2015. Microsoft COCO: Common Objects in Con- text. arXiv:1405.0312. Liu, H.; Li, C.; Wu, Q.; and Lee, Y. J. 2023. Visual instruc- tion tuning.arXiv:2304.08485. L ̈ uddecke, T.; and Ecker, A. S. 2022. Image Segmentation Using Text and Image Prompts. arXiv:2112.10003. Morales, P.; Klinghoffer, T.; and Lee, S. J. 2023.Fea- ture Forwarding for Efficient Single Image Dehazing. arXiv:1904.09059. OpenAI. 2023. GPT-4 Technical Report. arXiv:2303.08774. Pichai, S. 2023.An important next step on our AI journey. https://blog.google/technology/ai/bard-google-ai- search-updates/. Qi, X.; Huang, K.; Panda, A.; Henderson, P.; Wang, M.; and Mittal, P. 2023. Visual Adversarial Examples Jailbreak Aligned Large Language Models. In2nd AdvML Fron- tiers workshop at 40th International Conference on Machine Learning, volume 202. PMLR. Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I. 2021. Learning Transferable Visual Models From Natural Language Supervision. InIn- ternational Conference on Machine Learning, volume 139. PMLR. Salman, S.; Shams, M. M. B.; and Liu, X. 2024. Intriguing Equivalence Structures of the Embedding Space of Vision Transformers. arXiv:2401.15568. Salman, S.; Shams, M. M. B.; Liu, X.; and Zhu, L. 2024. Intriguing Differences Between Zero-Shot and System- atic Evaluations of Vision-Language Transformer Models. arXiv:2402.08473. Schlarmann, C.; and Hein, M. 2023.On the Ad- versarial Robustness of Multi-Modal Foundation Models. arXiv:2308.10741. Shao, R.; Shi, Z.; Yi, J.; Chen, P.-Y.; and Hsieh, C.-J. 2022. On the Adversarial Robustness of Vision Transformers. In ML Safety Workshop, 36th Conference on Neural Informa- tion Processing Systems (NeurIPS 2022). Szegedy, C.; Zaremba, W.; Sutskever, I.; Bruna, J.; Erhan, D.; Goodfellow, I.; and Fergus, R. 2014. Intriguing proper- ties of neural networks.CoRR. van Aken, B.; Risch, J.; Krestel, R.; and L ̈ oser, A. 2018. Challenges for Toxic Comment Classification: An In-Depth Error Analysis. arXiv:1809.07572. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017. At- tention Is All You Need. InAdvances in Neural Information Processing Systems. Zhu, D.; Chen, J.; Shen, X.; Li, X.; and Elhoseiny, M. 2023.MiniGPT-4: Enhancing Vision-Language Under- standing with Advanced Large Language Models.arXiv preprint arXiv:2304.10592. Zou, A.; Wang, Z.; Kolter, J. Z.; and Fredrikson, M. 2023. Universal and Transferable Adversarial Attacks on Aligned Language Models.arXiv preprint arXiv:2307.15043. Appendix Vision Transformers Very recently, several multi-modal models have been intro- duced. By using a shared embedding space among different modalities, such joint models have shown to have advan- tages. Vision transformers have been successful in various vision tasks due to their ability to treat an image as a se- quence of patches and utilize self-attention mechanisms. Transformer Encoder Block Linear Projection of flattened patches positional encoding MLP Head Transformer Encoder Block Class Dog Cat Bird .... Image Patches n blocks Figure 8: Vision Transformer (ViT) architecture (Dosovit- skiy et al. 2021) A collection of transformer blocks make up the Vision Transformer Architecture. Each transformer block com- prises two sub-layers: a multi-headed self-attention layer and a feed-forward layer. The self-attention layer computes attention weights for each pixel in the image based on its relationship with all other pixels, while the feed-forward layer applies a non-linear transformation to the self-attention layer’s output. The patch embedding layer separates the im- age into fixed-size patches before mapping each patch to a ImageTextMean PSNRMean SSIM ImageNet1,2,3-tokens43 dB0.980 ImageNetJigsaw toxic47 dB0.985 MS-COCO1,2,3-tokens46 dB0.986 MS-COCOJigsaw toxic45 dB0.982 Table 2: The average PSNR value and SSIM index between the original and embedding-aligned images to all the text embeddings in each of the datasets; the average is computed based on 800 examples for each dataset and the images and texts are strictly randomly chosen from the datasets with no postselection. high-dimensional vector representation. These patch embed- dings are then supplied into the transformer blocks to be pro- cessed further (Dosovitskiy et al. 2021). Additional Results Here we provide more details and additional information about the results we have included in the main text. +.04×= +.04×= Figure 9: Pixel differences between the original and corre- sponding embedding-aligned images in Fig. 1 (b) and (f); they are multiplied by 25 for visualization. Image Quality Evaluation.Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) are com- monly used metrics to quantify the differences between the original and modified images (Hor ́ e and Ziou 2010; Morales, Klinghoffer, and Lee 2023). PSNR effectively measures the detailed quality of an image, whereas SSIM provides an intuitive assessment of its structural integrity. We present the average PSNR and SSIM values between the original and manipulated (i.e., embedding-aligned) images across all the datasets under consideration in Table 2. These metrics indicate that the image quality does not significantly de- grade with minimal distortion. Due to resource and time constraints, we restricted the results to 800 examples for Ta- ble 2. We followed the approach by Szegedy et al. (Szegedy et al. 2014), where they used a smaller set (64 images) from ImageNet when calculating the average distortion of adver- sarial examples. More Results.In the main paper, the results are mostly gen- erated using the ImageNet and 1,2,3-tokens toxic dataset. To showcase the versatility of our framework across different vision and text datasets, the subsequent figures also present results obtained from additional datasets such as MS-COCO and Jigsaw toxic. Figure 14 shows the original outputs from the joint vision × text ImageBind model when an ImageNet example matches with all the 68 comments of the 1-token toxic dataset, there- fore getting 100% match success rate. It is clear that values are either very close to 1 or very close to 0, demonstrating that the classification results are stable. ballon -> "damn idiots"ballon -> "your beyond stupid"ballon -> "nonsense"ballon -> "poor snowflake"ballon -> "bully" 1.01.0087e-181.9061e-202.0126e-235.4837e-20 1.9376e-201.01.1445e-251.0169e-253.7303e-22 1.4496e-222.5056e-251.02.0181e-271.16e-21 1.6491e-232.2952e-241.1977e-241.07.0288e-25 1.1126e-186.3297e-192.8611e-178.8694e-231.0 vision x text hummingbird -> "bully"flamingo -> "bully"mushroom -> "bully"goldfish -> "bully"pelican -> "bully" 1.9846e-171.9265e-186.3261e-162.254e-211.0 2.4044e-191.1572e-191.5402e-177.2181e-231.0 1.2016e-173.2589e-182.9591e-163.3028e-211.0 5.7127e-186.0232e-193.8624e-172.0391e-221.0 6.7141e-191.5955e-181.1175e-164.7837e-221.0 vision x text Figure 10: Additional examples involving ImageNet and 1,2,3-tokens dataset. (top) Visually indistinguishable images have very different representations via embedding alignment with the corresponding texts and therefore very different classification outcomes. (bottom) Visually very different images have very similar embeddings, aligned and classified to a particular text. The examples are strictly randomly chosen. There is no postselection involved. ballon -> t1350ballon -> t6ballon -> t20009ballon -> t7020ballon -> t100602 1.08.3931e-242.9414e-212.3112e-231.0588e-23 1.6447e-241.05.1389e-225.5732e-253.2709e-25 1.3013e-202.8348e-201.05.179e-244.6858e-23 1.7716e-248.1671e-271.2559e-241.02.1426e-20 2.4776e-247.592e-261.8977e-235.3531e-201.0 vision x text hummingbird -> t100602flamingo -> t100602mushroom -> t100602goldfish -> t100602pelican -> t100602 1.9585e-208.7048e-249.7004e-141.1734e-261.0 1.9025e-197.9661e-244.523e-147.6604e-271.0 1.2284e-201.2407e-232.0851e-144.9291e-271.0 3.6575e-201.3402e-244.6727e-146.1279e-271.0 6.0804e-212.1571e-245.6463e-144.233e-271.0 vision x text Figure 11: More examples involving ImageNet and Jigsaw toxic dataset. (top) Visually indistinguishable images have very different representations via embedding alignment with the corresponding texts and therefore very different classification out- comes. (bottom) Visually very different images have very similar embeddings, aligned and classified to a particular text. giraffe -> "witch"giraffe -> "young and dumb"giraffe -> "gross evil"giraffe -> "clown"giraffe -> "awful comment" 1.01.1927e-223.7327e-203.8314e-171.7818e-22 9.7262e-261.09.0989e-253.2286e-257.7485e-22 1.1629e-224.0584e-241.02.7919e-221.7898e-22 5.3877e-212.9768e-267.3761e-231.06.3514e-26 3.1593e-247.22e-203.7954e-225.7248e-231.0 vision x text vase -> "awful comment" airplane -> "awful comment" cake -> "awful comment"clock -> "awful comment" bus -> "awful comment" 1.9585e-208.7048e-249.7004e-141.1734e-261.0 1.9025e-197.9661e-244.523e-147.6604e-271.0 1.2284e-201.2407e-232.0851e-144.9291e-271.0 3.6575e-201.3402e-244.6727e-146.1279e-271.0 6.0804e-212.1571e-245.6463e-144.233e-271.0 vision x text Figure 12: Additional examples involving MS-COCO and 1,2,3-tokens toxic dataset. (top) Visually indistinguishable images have very different representations via embedding alignment with the corresponding texts and therefore very different classifi- cation outcomes. (bottom) Visually very different images have very similar embeddings, aligned and classified to a particular text. Again the samples are randomly chosen. horse -> "naive" 012345 principal component 6 3 0 3 6 9 projected value (a) embedding projectionhorse -> "trash" 012345 principal component 6 3 0 3 6 9 projected value (b) embedding projectionhorse -> "nasty" 012345 principal component 6 3 0 3 6 9 projected value (c) embedding projection (d) Image (a) (b) (c) "naive""trash""nasty" 1.0 1.6401e-19 1.048e-18 6.2778e-19 1.0 2.6612e-15 1.9961e-16 5.7131e-15 1.0 vision x text train -> "naive" 012345 principal component 6 3 0 3 6 9 projected value (e) embedding projectiontrain -> "trash" 012345 principal component 6 3 0 3 6 9 projected value (f) embedding projectiontrain -> "nasty" 012345 principal component 6 3 0 3 6 9 projected value (g) embedding projection (h) Image (e) (f) (g) "naive""trash""nasty" 1.0 1.0219e-20 3.7767e-19 4.7651e-20 1.0 6.1847e-15 1.2138e-18 8.3552e-16 1.0 vision x text Figure 13: Same as Fig. 1, but shown while the proposed framework is applied on MS-COCO data examples and 1,2,3-tokens toxic comments. Again the samples are randomly chosen. (d) 1.0 2.3672728e-18 1.22242314e-17 3.7412264e-20 6.510062e-21 2.4181935e-18 4.2577153e-18 4.6259373e-19 2.3821952e-16 1.9946532e-18 6.1428153e-18 3.5659016e-20 2.1518446e-18 3.6921653e-23 5.8364814e-18 8.777897e-16 4.330217e-16 5.851934e-17 3.733509e-15 4.982623e-16 1.5942805e-16 1.0382484e-18 1.0979687e-19 3.1237739e-15 7.4216086e-20 9.2005165e-18 2.4658296e-12 1.11015405e-17 2.344426e-20 9.582661e-17 4.84685e-19 2.7053923e-15 2.5545014e-18 1.798785e-17 1.5795039e-16 9.13163e-19 4.2236408e-17 1.0695994e-14 1.4610572e-18 2.8870342e-16 7.787139e-18 1.4913891e-13 1.7022102e-18 8.910047e-15 1.6758327e-18 5.2495806e-18 8.0497784e-21 2.0336972e-20 7.550725e-17 1.7242399e-19 1.809248e-14 3.4992396e-20 3.6453154e-13 6.476512e-17 1.2795495e-18 5.251243e-18 9.431572e-18 1.3378695e-19 1.0338273e-08 3.8758954e-18 4.448262e-19 6.60757e-19 1.3602276e-14 1.9584475e-17 8.509951e-21 9.964077e-19 7.3634505e-15 8.050193e-15 1.9119511e-16 1.0 4.242371e-14 1.1153049e-21 5.432505e-18 8.5130435e-19 3.6332142e-18 1.6094461e-17 9.06096e-16 1.9199557e-22 6.5036086e-21 6.841929e-21 1.0179922e-18 1.07935e-22 1.5924287e-21 2.2082335e-18 2.4279322e-21 1.356241e-17 1.1563917e-17 8.8144387e-16 6.9177403e-16 6.749978e-13 3.074933e-23 6.736357e-17 5.812027e-23 1.5839698e-15 3.054405e-15 1.4734735e-23 9.081502e-19 3.099544e-17 6.6672324e-16 4.5180582e-17 2.313732e-18 2.701509e-19 7.004883e-19 2.3145497e-21 4.0184514e-17 1.0231529e-21 1.3054952e-15 2.3913326e-19 5.023556e-18 9.951206e-16 1.1138755e-15 3.8594087e-14 1.0461845e-17 4.105898e-18 1.2057975e-23 1.0996758e-17 3.8813105e-18 5.3935665e-16 1.1280603e-13 2.8863654e-14 7.014269e-19 1.0077257e-19 2.7990394e-18 1.2438606e-22 1.4992842e-14 4.7016097e-15 2.195214e-19 1.9650277e-23 1.8444691e-20 2.2218025e-19 3.6730594e-16 2.295742e-19 7.8522644e-16 5.2655046e-18 4.595514e-17 2.81329e-14 7.4384877e-16 2.8572704e-13 1.0 2.2659743e-22 1.751849e-17 2.2273195e-18 1.8248019e-14 3.0847337e-14 3.6862175e-13 6.612858e-21 3.2158667e-18 1.7009029e-21 1.7357783e-16 8.783541e-20 7.4255766e-19 3.9240036e-20 1.2478322e-17 1.5028838e-17 1.2761684e-15 1.888772e-12 1.8613812e-15 1.022627e-12 1.195904e-21 5.1865042e-17 4.8895943e-21 2.3650302e-14 5.2404305e-13 7.4380965e-22 1.3007592e-13 1.9161094e-12 6.455356e-15 1.3622305e-16 3.7624878e-16 1.0332081e-16 1.8406886e-19 2.1825782e-21 1.8350669e-14 1.5645934e-19 1.0308087e-12 1.2292604e-16 2.9223014e-17 7.368899e-14 9.11167e-16 3.485437e-13 6.3600036e-15 1.1835463e-13 1.9958135e-22 3.7539288e-16 1.1429293e-15 9.430687e-15 4.532052e-12 5.849109e-14 1.5636351e-16 1.7079452e-20 1.4216796e-19 9.797695e-21 1.7989335e-12 1.1405917e-13 6.2712336e-20 1.5405349e-21 1.4685329e-19 1.1698885e-18 2.8004334e-14 2.423965e-18 1.6212427e-14 7.033042e-14 3.4190491e-15 4.2018139e-13 4.4933376e-19 4.1910097e-22 8.46455e-23 1.0 2.0453128e-20 1.2411426e-18 5.2445234e-23 9.042406e-19 2.5805798e-21 2.3666341e-16 6.629358e-19 3.3606154e-22 1.351389e-21 1.3434611e-24 3.8817895e-23 6.911287e-21 3.8510492e-17 1.7812762e-18 1.31683e-21 1.2995048e-20 6.365064e-19 3.6388955e-21 1.9138278e-22 1.281273e-19 6.732021e-23 1.5518256e-19 3.5839903e-17 3.459511e-22 1.3826277e-23 5.7634823e-21 1.2236991e-21 3.7586496e-20 8.0910355e-20 2.624162e-19 8.790535e-15 3.564455e-23 1.1755261e-20 1.4754516e-18 7.025103e-22 2.4925733e-21 1.6844577e-20 1.4320052e-17 3.5617262e-21 3.79886e-16 2.2103623e-21 7.809221e-23 5.6226926e-22 5.22071e-23 1.8956957e-18 1.2672999e-23 2.255348e-16 2.863302e-22 5.6680176e-17 1.2601384e-20 4.3265216e-22 2.614722e-17 2.6761497e-22 8.1055124e-23 5.057546e-20 4.785727e-18 7.3424827e-19 3.7355347e-21 1.0780895e-17 3.5163762e-07 1.2805248e-20 1.8560062e-23 1.1462016e-19 9.190755e-18 1.6902774e-19 1.2993382e-17 5.739223e-19 1.7399058e-19 1.0 6.9860535e-16 8.1159865e-19 1.0069592e-15 2.7919708e-17 6.6693877e-18 8.314035e-18 2.4294832e-18 1.0129684e-17 3.8972986e-21 3.295895e-20 4.0442647e-17 9.874447e-17 4.629793e-16 5.415709e-16 1.271164e-15 8.7743487e-16 3.6980866e-15 5.4975876e-19 1.5774619e-17 6.786225e-21 2.894033e-18 3.0075897e-15 2.7557123e-22 2.0135745e-18 4.218104e-20 2.60385e-19 1.2981319e-13 1.8681087e-17 2.1960684e-19 3.3307863e-16 2.7694663e-21 9.11352e-18 3.7680835e-19 1.4195843e-16 1.37900785e-17 1.2522849e-19 2.3021376e-16 5.293767e-16 1.1466862e-13 1.5456466e-18 4.4893434e-20 7.2648466e-23 4.6139733e-18 1.0801853e-16 2.0556768e-20 7.091092e-15 3.584973e-18 5.242222e-16 1.4363151e-20 2.980447e-19 9.4886534e-20 2.0632997e-18 1.4979884e-19 6.0724678e-21 7.709472e-23 1.2856196e-13 4.6331146e-16 3.513555e-18 8.601574e-18 1.9333651e-18 5.4999814e-18 2.9673488e-16 1.2024165e-14 6.242714e-19 1.842479e-20 3.1758386e-21 6.8723025e-20 1.9858376e-19 1.0 1.10457636e-20 3.4347983e-17 7.29169e-17 1.849685e-18 1.0447e-19 3.4472762e-20 1.4633148e-20 4.7947237e-23 3.6586202e-20 5.273807e-18 2.6540532e-18 1.01199015e-19 2.0619059e-19 7.573481e-18 4.164007e-18 1.4783459e-20 2.4006165e-21 8.862876e-19 1.1672082e-20 1.0641467e-20 1.0803048e-16 7.994019e-22 3.1530688e-23 4.1017088e-15 5.5886466e-23 9.147458e-19 2.3527424e-18 1.2773006e-19 4.4461127e-17 5.5755755e-22 6.027409e-20 8.551189e-19 3.2024984e-18 4.9198325e-20 2.8459781e-21 1.8840728e-18 2.878462e-20 4.5317706e-14 1.30502395e-20 3.8005736e-22 6.2551657e-24 6.00843e-22 4.8201167e-18 2.6599872e-22 2.4673183e-18 5.519409e-23 3.1891144e-17 3.1560975e-18 6.9798825e-21 6.7274066e-21 8.83522e-21 4.0436547e-21 6.221e-22 5.9457895e-17 1.5074914e-21 8.380381e-21 1.4288686e-17 8.255911e-18 7.164074e-21 5.7721956e-21 1.751114e-17 7.189675e-18 1.15228075e-14 1.8061155e-16 1.5730835e-14 7.923855e-22 3.6067556e-18 6.9836977e-19 1.0 8.847625e-15 1.2711246e-16 1.2588104e-23 4.235215e-18 5.5876055e-22 1.6314467e-16 1.4657255e-16 3.905803e-18 3.9225877e-17 1.9528297e-18 6.267103e-18 8.8930825e-12 1.4487551e-15 2.6579072e-15 9.1821234e-15 1.5673978e-23 2.3011837e-15 5.5472133e-21 4.1958556e-15 1.9112036e-12 4.6225083e-22 1.4764165e-15 4.8761748e-17 5.7904046e-16 2.8532032e-16 3.932369e-18 1.4946273e-18 1.6795172e-20 7.659788e-22 2.3729362e-11 2.0341103e-18 4.8644564e-15 3.6398505e-15 8.368574e-18 4.234459e-15 2.8344995e-17 1.5040674e-14 1.1010749e-15 1.3423832e-12 9.977935e-24 4.877865e-18 1.12188695e-17 1.00309804e-16 4.3783883e-15 5.446208e-15 3.9031723e-15 2.5482893e-20 1.8126017e-18 1.5953342e-22 5.3337423e-15 7.665515e-15 1.0653495e-19 3.953121e-20 7.475143e-19 1.9002787e-18 1.4979914e-15 2.374016e-20 1.1273064e-15 3.7272046e-15 2.1955859e-16 1.07056155e-14 1.6397673e-19 7.459281e-18 1.2469722e-15 1.5591062e-19 9.968839e-17 1.2680201e-16 1.02317894e-16 1.0 8.597575e-18 1.04686615e-19 7.105828e-21 3.8908942e-22 2.4554514e-17 9.910962e-20 1.724678e-22 1.8482797e-20 4.3103325e-20 2.2105176e-18 5.185713e-17 1.7165846e-15 4.19689e-18 3.3475624e-16 2.6955137e-22 2.8466883e-18 1.4217192e-24 3.9492984e-17 2.1729424e-13 3.3549598e-26 2.6114936e-16 5.210701e-16 1.6714992e-16 4.9679525e-16 2.3737668e-17 8.319325e-20 9.906771e-19 2.0077978e-19 7.1042337e-16 4.018394e-19 6.3583513e-16 7.355855e-19 6.6213244e-18 1.6085548e-17 2.6508051e-18 8.936194e-14 3.702856e-16 1.9144751e-17 2.7300746e-24 3.0940244e-19 1.5968292e-18 6.874511e-18 4.3085703e-14 2.4099791e-18 7.917073e-18 1.160076e-18 1.3439302e-20 2.3763168e-21 1.1417484e-15 1.0353746e-15 1.2598294e-22 5.784441e-17 4.5590664e-20 2.247347e-21 2.5818565e-16 1.3558482e-16 8.222258e-17 5.0241357e-16 1.3276229e-15 3.3546939e-15 5.774302e-16 1.0917355e-15 4.5203213e-15 2.319896e-22 1.4446752e-17 1.4424575e-15 1.5050353e-17 4.88804e-16 1.0 2.5854696e-19 5.023228e-19 1.4585673e-19 2.3902943e-18 1.2240878e-19 6.600138e-19 1.0611253e-15 6.3886385e-18 1.4352061e-16 4.243095e-16 1.9189026e-13 3.262167e-16 7.702738e-16 1.07804317e-19 1.8878747e-16 3.9073343e-20 6.647246e-16 3.4992083e-15 3.6648723e-23 1.0981955e-18 1.781244e-16 5.395372e-18 1.6662734e-15 4.4912003e-17 1.7389245e-18 1.7334673e-18 5.7799824e-18 5.407936e-16 1.3719311e-15 1.6077104e-16 2.9682642e-17 2.9210948e-19 1.4611857e-16 1.8531749e-17 2.0398337e-14 6.514153e-16 9.0110105e-17 3.3176372e-23 1.0699813e-17 6.893212e-17 1.09715756e-17 9.170917e-16 1.5947185e-16 4.1641523e-17 1.20115525e-17 2.3854083e-21 3.47384e-18 3.4360074e-15 5.3816408e-18 9.672294e-20 3.5681857e-21 1.7461818e-20 1.5118282e-13 9.831684e-15 2.4393553e-17 8.668255e-16 7.351995e-16 1.6655744e-15 8.468129e-15 2.6207506e-19 3.9598312e-23 7.4802366e-22 4.277859e-16 4.0966027e-19 1.2689105e-16 1.3418022e-23 8.122668e-21 2.7260632e-20 1.0 9.8243805e-21 5.005171e-19 1.2559864e-21 5.599527e-24 1.6847581e-22 2.8241323e-21 1.6511419e-18 9.966737e-19 3.4832826e-21 8.5163524e-20 2.937958e-19 5.9232266e-22 5.526201e-20 6.98516e-20 5.2990243e-22 5.0851085e-22 5.1299674e-19 3.4278946e-20 2.2095603e-22 4.9167417e-18 2.936428e-23 9.6352785e-21 1.0736658e-16 6.185343e-21 1.1346396e-17 1.4966969e-20 9.175525e-22 2.331551e-17 4.2296246e-20 3.361954e-19 2.4256272e-21 1.716141e-19 8.048327e-23 5.48379e-17 2.411519e-22 4.4297497e-21 4.6766425e-21 4.4567287e-23 1.07128e-17 4.8402438e-23 3.2462225e-19 1.6839935e-22 2.7506255e-17 6.7226416e-19 2.7758838e-23 2.1797166e-13 3.9379904e-22 2.4404388e-22 1.3564727e-19 5.635647e-20 2.1958063e-21 2.0580544e-20 1.1075437e-18 2.450539e-16 1.4504801e-23 6.537533e-22 8.674923e-19 3.5320274e-19 8.674866e-17 6.3340745e-22 2.4495175e-19 1.9531875e-18 2.0316145e-17 6.5790013e-18 1.2453303e-18 4.3449533e-19 1.2419573e-18 1.6388544e-19 1.0 7.834514e-18 1.5140333e-20 2.2959833e-23 5.8080907e-20 3.4082634e-17 4.270689e-14 5.892106e-14 1.8919897e-18 1.3249957e-17 7.727364e-18 3.820918e-18 1.2148552e-17 2.950426e-18 8.814315e-21 8.047142e-20 3.9436118e-15 6.196739e-17 6.822698e-20 3.0197024e-18 1.1792287e-21 8.376974e-18 5.416465e-19 7.526191e-18 2.6268173e-18 1.1368951e-20 3.5874414e-18 9.336318e-18 6.751529e-19 3.484019e-18 1.0182301e-16 1.1175427e-16 3.9311884e-16 5.301065e-15 3.0875952e-21 4.4565826e-20 9.422402e-18 1.2221125e-22 1.1747016e-16 4.768299e-21 7.734677e-18 4.1980822e-22 2.372227e-10 2.8307388e-17 7.867394e-19 2.664743e-20 3.8659705e-19 1.869881e-20 7.043924e-19 1.0682144e-19 4.8406594e-19 6.728514e-20 4.1866284e-16 3.7945987e-17 7.906407e-21 2.9654996e-19 1.7882358e-16 1.6342783e-17 4.8874737e-19 2.5688022e-22 4.468869e-22 6.785086e-21 2.9802523e-20 2.4439467e-18 1.2323368e-22 3.595717e-22 1.1482222e-21 7.8168077e-19 1.6958261e-18 1.0 2.258945e-18 9.916922e-23 1.8390284e-23 4.8820907e-18 1.1840942e-18 3.8502395e-18 3.3067474e-22 1.07051505e-19 5.3858332e-17 3.6621493e-22 5.972682e-17 1.376231e-20 4.4907454e-21 4.462172e-23 2.873055e-19 1.585026e-17 7.737727e-23 5.825354e-20 9.32085e-23 5.647462e-20 6.05572e-21 5.058282e-19 1.9936947e-18 4.3541343e-24 5.879362e-22 6.1128415e-20 3.522342e-20 4.8223192e-20 7.566077e-21 1.5346396e-20 4.7716108e-21 6.866121e-19 4.443473e-22 3.5482176e-22 4.115661e-21 1.0705513e-20 3.9293382e-19 8.5897963e-20 1.0203861e-20 4.5933966e-23 9.5699854e-17 4.777464e-18 8.385626e-21 2.8898506e-19 1.1222224e-22 2.9730698e-24 3.4620117e-19 5.9473426e-21 2.4888392e-21 4.935999e-20 4.865934e-18 5.250675e-20 8.361951e-22 3.4114172e-21 1.3472256e-20 3.7106516e-19 4.7185125e-17 1.3623702e-17 2.4616895e-16 1.2997328e-20 1.3548707e-17 8.198231e-18 7.2767396e-17 9.44022e-16 2.2145602e-18 2.5359831e-20 7.7910526e-20 7.425301e-17 1.0 2.4956883e-18 1.524814e-20 5.915581e-19 9.33642e-19 2.258971e-18 2.1290335e-16 4.524688e-14 1.1686636e-15 2.0084453e-15 2.9232684e-19 4.5760637e-16 6.967782e-20 6.189799e-15 1.1211829e-14 8.970941e-19 2.0756496e-15 3.3809566e-15 3.340597e-16 2.4025569e-15 1.10002415e-17 7.525474e-18 1.585928e-16 1.2158043e-19 4.514259e-17 1.6989739e-17 1.7637663e-15 7.4981156e-16 6.504692e-16 2.2267422e-16 4.6047265e-17 1.0307492e-13 9.98639e-16 2.2650554e-17 1.3499862e-23 4.642522e-18 3.0746962e-17 8.584147e-16 3.0827222e-14 3.6185442e-15 2.217827e-15 3.5284105e-18 4.4808275e-18 2.6241018e-19 6.198846e-16 6.0168866e-17 3.5904496e-17 6.448789e-21 7.904576e-19 1.0804521e-18 2.5129737e-15 5.889443e-17 2.04357e-16 7.0173605e-15 2.9536482e-17 1.517107e-13 7.15149e-20 5.5447893e-18 1.2616699e-17 2.9406207e-23 4.3004805e-19 7.4081205e-19 2.898428e-12 1.16165055e-14 3.5707667e-17 2.2839906e-22 7.699311e-21 3.4586832e-20 1.743576e-16 1.0 4.4178905e-19 3.023272e-20 1.023782e-20 1.6800754e-19 6.893712e-17 4.2732055e-18 3.1254378e-18 1.2808602e-15 1.1183607e-19 7.461874e-17 9.199853e-25 1.0420661e-17 3.5625257e-15 1.7120252e-24 2.299251e-17 9.662185e-21 1.04749885e-16 4.1393977e-17 3.556344e-17 1.7755757e-20 3.341337e-23 4.983048e-21 6.9064735e-14 9.047782e-20 9.8586774e-17 2.6498994e-15 2.736665e-16 2.0464147e-17 1.974892e-18 1.3816406e-17 1.6188054e-17 5.3647187e-15 1.5654671e-24 1.0631285e-12 1.1183644e-20 9.230367e-17 3.392157e-15 3.7063872e-17 2.36991e-19 1.9608459e-19 2.7590509e-19 1.9300139e-24 9.897763e-15 1.7606579e-18 1.3816627e-23 4.293982e-22 2.4708372e-21 7.642484e-21 2.9452708e-14 1.342705e-20 5.3850113e-17 3.0229801e-15 2.9000671e-18 1.33867105e-17 8.300152e-19 1.5289538e-22 1.1716648e-20 5.0436276e-23 1.5640309e-21 2.9178093e-19 3.849854e-20 1.6923855e-21 5.4053578e-20 8.606708e-23 1.0544124e-20 9.024227e-23 1.0534429e-21 7.0409547e-23 1.0 1.0228379e-19 9.80973e-15 4.952063e-22 1.4519065e-18 1.5642198e-16 5.7337026e-21 2.2372804e-20 1.078976e-21 6.6457843e-21 1.9096316e-10 7.061859e-22 3.3326054e-15 3.5799778e-22 1.4923844e-23 9.7335144e-20 5.9888228e-21 2.7195818e-20 2.0949788e-16 1.4057054e-20 1.7741943e-20 8.205451e-20 3.2924328e-20 2.384268e-16 8.220516e-21 7.204838e-18 1.6289575e-18 3.5891197e-16 2.5510709e-20 1.2829995e-19 8.880503e-20 3.0489958e-20 7.0453825e-17 3.7454639e-22 7.159273e-19 1.0667228e-21 3.0783092e-18 6.2234957e-21 1.667953e-14 1.3613025e-21 1.6694075e-18 4.884338e-21 1.0439937e-21 4.7340215e-23 3.6856238e-22 7.054438e-19 7.984754e-22 1.836969e-21 6.1086785e-17 3.7151973e-21 6.4618897e-21 3.369788e-18 6.307059e-18 9.824165e-17 3.7632664e-14 3.09169e-18 4.0113468e-20 3.8932423e-21 3.8751695e-19 8.3629977e-16 8.090052e-19 2.1006173e-19 9.5659335e-17 3.5899096e-20 2.3753156e-17 1.9738162e-17 1.1473691e-20 5.1739995e-23 4.3577202e-18 1.0 4.811451e-16 2.8173265e-17 5.474507e-18 9.911841e-18 8.218675e-19 3.7785336e-19 5.143777e-21 2.8998556e-19 8.005353e-19 2.0539927e-19 2.541201e-14 8.157944e-21 6.6643167e-21 6.095012e-18 6.693872e-21 1.638056e-17 2.5550569e-18 5.110472e-20 1.6971335e-18 2.643422e-19 2.0033203e-17 1.2501812e-17 2.9754485e-19 2.9510115e-18 9.200788e-20 3.5388174e-16 1.0018044e-18 3.924835e-16 2.7858652e-16 9.649477e-21 2.7639688e-20 4.6290045e-21 4.8342586e-18 4.6418975e-20 7.330815e-18 6.2516004e-20 8.03762e-15 8.0101226e-18 6.147266e-19 7.9919845e-20 4.7860117e-21 1.5706692e-19 1.03454445e-17 3.427779e-19 2.2314929e-20 3.4026478e-19 8.4552533e-16 5.9769846e-18 8.717971e-22 2.6226322e-18 2.6898918e-17 1.2525012e-15 4.3477534e-20 1.4643903e-23 1.5832408e-20 2.771296e-20 4.7888063e-21 6.720562e-20 1.3654527e-21 1.08631546e-20 2.8326992e-21 1.6116108e-22 2.4141196e-18 2.9972827e-22 2.439722e-22 8.503703e-25 2.4936316e-19 1.4766775e-20 1.0 1.2085445e-19 1.0449897e-21 3.6628292e-17 3.0016466e-19 6.5674283e-22 3.3845226e-20 5.152679e-19 7.71525e-20 4.2753472e-23 3.716547e-20 1.5730752e-17 1.0809194e-23 2.224796e-19 1.2846823e-22 7.335116e-20 7.3719196e-20 9.698849e-22 2.985085e-22 1.5738275e-22 3.2446773e-21 2.0238374e-20 1.411788e-21 3.9744542e-21 3.1996485e-21 3.845154e-20 3.6087596e-21 7.159024e-20 9.4416995e-23 1.721522e-23 2.031828e-13 7.6052835e-22 2.828975e-20 1.945976e-23 1.9309751e-20 2.5393287e-23 3.4537444e-17 7.9737964e-22 1.4376895e-22 1.0457834e-21 8.8140376e-23 1.2975405e-24 3.751972e-21 4.2015652e-19 5.0102882e-20 2.4315842e-21 1.594438e-17 2.690102e-20 1.5472209e-19 6.745983e-22 4.573631e-19 1.2840159e-20 4.803968e-16 6.707068e-19 6.120828e-18 1.27487764e-17 1.6633575e-16 3.1820887e-18 1.7560373e-17 2.1859742e-17 8.6808575e-17 1.7113326e-17 3.4149058e-14 2.013682e-18 1.223033e-19 7.695992e-21 1.7307037e-20 4.598074e-17 1.1762266e-15 1.0 2.653752e-16 1.5176562e-16 5.714938e-16 1.4015419e-17 1.7719913e-18 4.5437597e-17 1.5484955e-20 3.871682e-19 9.385329e-14 3.3124878e-17 1.3264434e-15 2.8188346e-15 2.4006007e-19 2.6414331e-14 7.1247885e-17 1.7751304e-18 3.8641167e-17 5.0688334e-17 8.5521634e-16 6.58639e-16 3.4358613e-16 4.2108092e-18 3.1591568e-18 2.683791e-17 8.739736e-17 1.03825504e-13 4.511957e-19 3.6604823e-17 1.4367158e-19 1.7255559e-19 9.620204e-17 3.1067503e-19 9.9063986e-18 6.913452e-20 3.962811e-14 1.5410972e-16 9.2281704e-20 8.050588e-18 9.424091e-18 4.8427124e-18 4.6014784e-17 1.4148141e-19 1.1841225e-16 6.906079e-18 4.8465823e-15 1.8392718e-17 3.3600325e-18 1.5972616e-18 3.6361966e-17 2.5663883e-16 4.65995e-16 1.3479827e-18 2.1840976e-18 5.860948e-21 1.5414433e-17 2.6367364e-18 3.918152e-14 8.306839e-18 9.20497e-19 2.168008e-20 2.9143401e-18 7.639128e-22 7.326842e-19 1.7878017e-21 7.132426e-18 5.016602e-19 9.8271315e-18 5.975961e-18 1.0 7.232707e-18 5.0891185e-17 2.3926444e-21 1.8096989e-19 1.7416265e-18 3.0799755e-19 2.2116718e-19 9.992412e-16 1.4118757e-19 1.7470096e-18 6.2349587e-18 4.318677e-20 2.803775e-17 1.435879e-17 1.0799606e-20 3.5905393e-20 1.3764425e-18 8.068279e-11 1.4551327e-19 7.476605e-17 4.4583066e-17 1.5334588e-20 2.9016694e-18 2.9552343e-19 9.058848e-17 1.1126289e-19 1.6846637e-12 1.4681737e-20 3.7793515e-21 1.0495278e-14 1.3330451e-19 1.5270457e-17 2.3962351e-20 1.785014e-14 1.7501524e-17 8.5280835e-21 3.4472762e-20 1.6853841e-19 1.1099116e-19 4.8895833e-18 5.5291717e-21 6.002166e-18 8.263126e-19 2.1635996e-17 2.5763913e-19 8.06447e-21 1.1604522e-18 1.6675196e-15 5.5049924e-18 9.695582e-17 2.048063e-18 7.9811547e-16 5.5871637e-20 5.262395e-17 2.607232e-17 6.953954e-18 2.3173207e-16 1.2228717e-16 2.1584026e-20 1.6684844e-18 1.3007552e-19 1.4144539e-17 4.3600557e-22 1.5789075e-16 1.3843253e-18 1.4386014e-13 1.2621169e-18 5.2398e-17 1.0 6.278667e-16 1.7637529e-15 3.5272514e-19 1.1046719e-15 2.4497042e-19 5.41496e-18 7.825893e-15 2.0506028e-19 1.9671336e-17 2.9106892e-15 2.9061804e-18 9.854019e-16 1.6197892e-15 1.2635332e-18 3.2127385e-19 3.4856428e-17 4.5599512e-17 4.488922e-17 4.5092572e-18 3.0585907e-15 6.0717133e-18 4.970017e-15 6.5953087e-18 9.22478e-15 2.786184e-16 4.97654e-18 4.9703143e-19 2.8197338e-18 2.6689513e-15 7.035519e-19 2.277203e-16 3.923424e-18 2.2093513e-13 6.077248e-19 1.8649737e-19 8.2330544e-17 1.5653248e-15 1.2657564e-19 5.9070764e-20 2.3598433e-18 8.965262e-19 3.1578902e-19 5.3701647e-15 5.644324e-17 1.6357028e-16 3.5068404e-14 3.2582867e-16 3.991144e-15 1.5972703e-15 2.2293764e-18 2.1269716e-16 4.889301e-19 1.4832011e-16 4.6017087e-16 2.2156525e-16 2.8613236e-16 1.7076435e-15 1.0866973e-19 7.897557e-18 1.9938715e-16 3.1780525e-16 2.5054724e-20 4.5546684e-20 4.4676e-18 3.5558708e-16 8.819483e-16 1.946406e-16 7.257952e-15 1.0 7.60675e-18 5.1683907e-18 2.3109388e-08 2.1029595e-18 2.7139377e-17 2.1049933e-15 1.490687e-18 4.581037e-18 1.0016532e-15 5.44352e-18 6.5157364e-18 1.081888e-17 1.6080102e-17 4.855446e-15 2.6111793e-20 8.9521986e-17 1.5512649e-17 6.1967416e-16 4.32112e-18 4.9131954e-19 1.3961626e-16 6.369886e-15 4.272113e-11 3.8910338e-19 1.9989274e-17 2.1411229e-21 6.4398576e-16 7.667065e-15 2.7532097e-16 1.482953e-15 1.7714495e-19 2.6704498e-14 1.3291171e-16 1.200993e-20 1.7429749e-19 3.365462e-16 9.706932e-19 1.06971255e-16 1.1472902e-20 7.6198515e-17 1.4702145e-19 2.3953377e-14 3.8569418e-18 9.923142e-13 7.476662e-17 3.3784796e-16 1.0612022e-13 3.0601849e-18 4.2557917e-14 3.274374e-15 1.5936546e-23 1.0659497e-14 5.383281e-19 7.695568e-17 5.3107976e-16 6.434108e-17 8.4454e-21 7.240188e-18 1.2291871e-21 3.3050534e-17 8.2746664e-20 1.1184844e-19 1.4408048e-19 3.7041765e-20 1.6348528e-18 3.809155e-20 1.6960115e-13 3.2267542e-18 1.0 1.9881029e-20 5.8828634e-18 2.697539e-23 4.1074062e-16 5.54749e-13 5.064916e-22 2.7376883e-16 7.083755e-18 2.5747946e-16 1.09770236e-16 2.9091545e-17 1.6933172e-19 3.6305954e-18 3.2693732e-20 3.787178e-15 1.3484765e-18 4.741525e-14 5.5649224e-17 5.180908e-17 6.19279e-16 2.0107165e-13 2.2083726e-14 5.995925e-16 3.3983724e-16 7.9593006e-22 7.3180436e-19 1.756673e-18 8.6794335e-17 3.045972e-13 1.117856e-14 7.345203e-18 1.6327998e-21 2.2218193e-19 1.539183e-22 2.8960865e-11 6.554055e-17 5.7202644e-22 1.6319778e-21 1.2726115e-20 4.658859e-19 4.43981e-16 6.1783195e-19 8.9485e-16 6.366482e-16 4.8792635e-17 1.8504503e-13 4.291942e-19 3.1812432e-23 1.5573688e-21 2.2114502e-21 1.7467815e-20 9.0181486e-20 2.4767905e-22 4.4580596e-21 6.1771166e-20 5.9677507e-21 3.485749e-17 6.946299e-16 5.1711632e-21 9.353262e-23 1.5035867e-20 6.8410454e-20 3.949888e-16 1.252826e-17 3.2731932e-19 2.1025905e-18 3.6555965e-18 1.0793141e-20 1.0 6.2918476e-18 1.1993527e-22 2.9116894e-22 1.3283541e-20 1.1758923e-19 5.5496262e-21 1.0270448e-19 1.6771747e-19 4.780725e-19 5.5815847e-19 9.752683e-22 8.610292e-22 6.764802e-20 2.838471e-20 1.8332108e-20 6.840853e-23 7.0069405e-19 4.1779067e-21 3.1693734e-21 1.04570473e-19 8.223406e-20 9.507194e-22 1.0475921e-21 4.3627433e-18 1.6375492e-18 2.4311068e-17 6.429278e-16 8.480913e-21 4.607843e-22 1.1511643e-15 3.535e-17 1.8114306e-21 9.437748e-21 2.6669802e-20 2.0173377e-22 6.1377702e-21 1.4845906e-20 2.1241528e-21 5.194161e-19 4.4319962e-14 8.874395e-22 3.3914208e-22 7.538062e-17 4.196311e-19 7.41373e-21 1.4766104e-12 1.7366102e-19 9.647352e-19 4.0494145e-20 2.8101553e-18 1.2557111e-17 2.1296662e-17 2.78064e-17 1.3468628e-17 2.0883291e-19 3.0124307e-18 8.8489527e-20 1.1819872e-17 1.6608995e-21 1.3383041e-18 4.33704e-18 3.0473056e-15 2.3119336e-17 1.447009e-16 1.8793047e-14 3.3468737e-08 2.6373411e-17 1.0728621e-18 1.0 7.977542e-21 1.7103594e-18 9.816244e-15 1.1632438e-19 1.0653245e-20 8.4792485e-17 8.010673e-18 1.0966095e-17 4.9802808e-18 1.249718e-18 4.567409e-18 3.2141463e-20 4.832408e-17 6.7477027e-16 1.0283794e-16 1.8495822e-15 7.8352574e-19 2.303597e-15 1.512847e-16 3.0189726e-13 6.490111e-18 6.3179636e-19 7.026691e-20 1.3870535e-17 3.3836515e-16 3.3139378e-19 1.13682e-15 5.743885e-20 1.7100757e-14 3.4026627e-18 1.3953737e-19 8.5831517e-19 6.5959667e-17 6.448986e-21 7.925474e-18 8.3301224e-20 5.4758258e-17 7.8492836e-20 8.3909625e-15 9.7007896e-18 6.7777345e-15 1.5591916e-16 1.6243681e-15 3.83194e-15 1.5599859e-20 8.860858e-24 1.7320822e-22 2.4526356e-23 1.0891935e-23 2.8898052e-20 2.7871764e-21 4.4539415e-23 2.8893405e-21 1.1663346e-22 2.0683558e-21 1.0450934e-21 6.2883583e-21 1.4285242e-26 1.649644e-10 3.6760233e-20 1.1070216e-15 1.6027707e-21 5.4288536e-20 8.191096e-20 3.162869e-19 2.0164596e-23 8.559457e-24 3.206255e-23 1.0 4.4452508e-23 7.7111643e-19 2.7266808e-17 1.4260319e-24 1.8336506e-21 1.8117804e-24 7.8260444e-22 1.3361048e-19 6.335736e-24 1.12594725e-17 1.6039573e-21 5.3791125e-23 8.870215e-19 3.3441808e-23 5.126379e-22 6.800968e-22 1.3917416e-19 1.6926881e-22 1.052058e-20 3.662151e-21 2.294365e-22 1.00787894e-20 3.1353246e-26 1.656705e-20 2.3152876e-24 3.649182e-21 1.9752854e-24 1.8882565e-16 3.87305e-22 3.043345e-22 1.0565089e-22 1.9916284e-24 1.3389434e-24 4.0526373e-16 1.9108972e-20 1.9080625e-21 4.85968e-19 4.7290626e-21 4.288039e-22 4.936594e-23 1.6463578e-23 2.7362107e-20 9.499881e-18 1.0797075e-16 6.1691434e-17 7.306614e-17 1.6417397e-18 7.9977605e-17 4.0605246e-18 3.1045039e-16 4.8723864e-15 5.192089e-16 2.2071459e-19 9.421566e-20 9.916771e-23 9.021883e-16 5.1472538e-20 1.9342189e-20 1.394342e-18 8.658913e-21 8.807803e-19 1.5895794e-17 1.0902497e-15 1.6389202e-15 2.1947256e-14 1.1706226e-22 2.8948406e-17 1.4623654e-21 1.0 1.0885926e-13 6.4737574e-21 1.8124237e-16 1.3137814e-18 2.1269877e-16 8.203925e-18 1.2524928e-15 2.99282e-19 2.239006e-19 7.082802e-20 4.409201e-16 2.027213e-21 1.0880726e-15 4.272002e-17 2.6568618e-16 4.1232584e-16 1.1123129e-15 1.4768285e-14 8.209918e-15 9.91551e-18 1.2033568e-24 8.667977e-19 2.0988103e-16 2.0451443e-17 2.6233378e-14 1.687941e-16 5.2895457e-17 5.9434245e-19 2.1702907e-21 5.0473723e-19 6.428766e-15 2.940341e-15 3.2925415e-19 4.2463713e-21 1.5837733e-19 4.2881848e-16 1.5980503e-15 8.601065e-14 3.487333e-16 4.445201e-15 4.2886263e-16 2.8222532e-15 4.4568196e-13 1.2851757e-17 3.9624294e-16 5.847555e-19 2.082365e-17 7.735419e-17 1.9523923e-16 8.6189276e-16 5.32707e-16 2.7398155e-19 3.1679922e-17 2.787805e-20 2.1186216e-17 3.735451e-20 1.20659445e-14 1.2310772e-15 3.2346802e-17 7.603649e-17 1.3328495e-15 4.7325776e-15 6.511888e-17 5.3668677e-15 1.6307207e-21 2.0976656e-16 2.2922525e-18 1.4438008e-15 1.0 1.6093033e-19 1.9930255e-18 5.579656e-16 5.4971945e-17 8.4282356e-16 7.185139e-16 3.022328e-19 7.975668e-18 2.0921575e-18 1.1176826e-15 2.948611e-13 8.255769e-15 1.23758984e-14 5.3823237e-16 6.9477616e-13 1.6012723e-15 1.08874207e-13 2.0588214e-15 2.6608162e-17 1.1752056e-20 2.4744347e-18 6.187054e-17 1.0386886e-17 2.1508243e-13 4.2778076e-17 1.7381474e-12 2.744083e-19 6.692959e-19 6.044091e-20 1.9181972e-16 5.7593377e-18 2.0121397e-17 3.435755e-17 1.7825471e-18 1.4289544e-19 3.991714e-14 3.5713005e-15 6.0588486e-18 7.468998e-16 8.686299e-15 2.5908952e-13 1.7156057e-19 6.4604257e-24 7.676487e-23 3.9151333e-22 3.6840597e-24 2.5015557e-21 6.9408974e-23 9.299925e-24 5.8883206e-25 3.2430488e-22 5.385625e-18 2.8926063e-20 3.7675353e-20 1.71092e-26 1.9906345e-23 1.0177183e-22 1.2087782e-14 1.00273684e-19 2.0039086e-21 3.5589304e-20 1.770888e-20 1.0393038e-23 4.1882286e-22 3.3246956e-22 3.1752136e-18 7.5048995e-23 1.935171e-20 1.0 1.6039925e-23 4.5381204e-21 2.2858174e-25 4.7180706e-21 4.811016e-21 5.8567563e-25 1.7371063e-20 1.6340467e-20 2.021e-22 1.89964e-19 1.852513e-22 2.2871132e-21 7.4165996e-20 1.3152777e-21 1.806799e-22 2.1706799e-21 1.4768372e-25 1.4452008e-22 5.958953e-19 7.17902e-26 1.7455349e-21 5.843949e-24 4.2658985e-23 5.4020253e-25 4.8569233e-17 1.1756528e-22 1.0607568e-23 2.4629406e-22 8.492519e-25 6.1540287e-25 5.03386e-17 1.2041623e-21 3.2175765e-23 1.5870442e-20 4.277801e-20 8.975511e-22 2.4383716e-23 1.30093616e-23 3.1977953e-20 8.4420375e-22 2.7404128e-18 1.8327502e-19 1.16298e-15 2.99926e-23 1.3302526e-17 8.0706293e-20 4.1805933e-17 5.869151e-16 6.529923e-17 5.25883e-22 5.5131104e-18 2.115363e-21 3.6069227e-17 1.0565468e-19 1.6884897e-22 1.4943578e-20 1.0545968e-21 2.1391281e-16 9.988557e-17 6.280994e-15 3.0007452e-17 3.495153e-15 2.640225e-21 5.787194e-19 1.1183073e-23 9.736017e-17 2.0937265e-14 4.795526e-24 1.0 1.0235571e-17 2.4461502e-16 5.216861e-18 1.0095231e-18 6.4412017e-18 2.3278048e-21 1.2803056e-19 3.1484015e-14 6.29546e-22 2.1559242e-14 2.787428e-16 6.383161e-17 1.1967178e-17 2.0419803e-16 5.4412004e-16 2.1997656e-19 1.8342367e-15 8.414736e-25 1.3492231e-17 7.0786457e-19 1.5029492e-14 3.7130544e-16 2.1844986e-17 1.8335482e-18 2.2071699e-20 1.7474095e-18 1.1021466e-22 1.299632e-15 1.2520467e-18 3.4277773e-20 2.3935113e-21 1.3271543e-19 1.4237915e-18 4.1192528e-16 1.864206e-18 2.0900537e-16 4.5391062e-14 3.0538558e-16 1.6506766e-15 1.3105898e-15 1.0608607e-18 4.541254e-14 1.1564864e-21 6.2701543e-21 2.1050336e-15 1.1876725e-18 8.547778e-16 7.804428e-15 6.5393177e-18 9.056386e-19 2.3357671e-20 2.7676977e-17 2.5688009e-23 4.9942617e-19 5.017272e-19 4.3148327e-16 4.0503652e-16 2.8325318e-17 1.575727e-13 2.8407126e-17 2.0666171e-17 1.05574055e-20 1.2405852e-16 3.9354547e-21 1.6009339e-18 3.183261e-14 6.8720925e-20 1.0614152e-18 1.0 1.8564003e-18 1.1142957e-16 1.1009284e-14 6.968108e-18 1.1659717e-19 2.349095e-16 3.5569354e-16 3.328427e-14 3.968052e-18 6.5739584e-18 4.382223e-19 2.0860614e-16 1.8027605e-19 2.1418087e-14 6.8668845e-18 1.6885383e-18 5.612682e-23 1.1164352e-17 4.177537e-15 7.322822e-18 1.352037e-16 1.637643e-18 7.696604e-15 4.732896e-18 6.267168e-20 1.197071e-16 3.7633348e-16 8.723473e-19 7.044623e-19 4.045673e-15 8.328756e-20 4.4008973e-21 2.4296338e-16 3.6394967e-18 1.8841163e-19 5.011655e-16 3.2302446e-15 7.8798886e-17 2.7137498e-18 6.4484857e-16 1.0574839e-16 4.464303e-22 4.2509902e-21 1.2608157e-21 1.2318129e-18 7.489797e-16 8.044817e-18 4.475845e-23 2.106953e-23 2.0858964e-23 4.2068276e-18 1.8695397e-19 2.1652899e-19 8.631681e-19 3.578123e-21 2.8884824e-20 7.4214986e-19 1.21096e-16 9.344686e-19 1.9908078e-17 1.34309486e-20 6.738002e-17 6.621582e-24 2.0694975e-16 3.708355e-16 1.130461e-26 9.1530175e-17 7.283679e-18 1.0 6.4961054e-18 1.508558e-18 4.509909e-19 7.137526e-23 1.2073958e-21 1.05197766e-16 5.76386e-20 1.642418e-16 1.7316648e-16 2.2486367e-18 2.4998227e-17 2.121079e-18 1.3661647e-16 5.2434967e-18 1.6727741e-17 2.1801357e-25 3.1313723e-16 1.614955e-19 1.16756634e-17 3.4672919e-15 1.8708133e-15 1.6267537e-17 1.1230746e-22 2.8673244e-22 5.997874e-22 1.1146023e-15 2.8337765e-16 2.2181226e-23 4.281521e-21 4.2624316e-24 4.1661e-20 3.442802e-16 1.27124275e-20 4.114804e-18 8.126749e-16 3.8520946e-16 2.1914248e-16 4.1372034e-17 6.278223e-20 4.1439122e-19 8.619953e-22 5.4545603e-17 1.2442524e-18 7.53334e-19 1.9030435e-17 1.2000148e-17 3.7093073e-21 1.3322731e-18 2.484182e-21 1.7258136e-18 1.408367e-21 7.245177e-21 1.7629361e-18 1.3445631e-17 3.7158296e-17 6.0552175e-19 1.2762075e-15 8.89641e-19 4.577279e-19 2.5092302e-21 1.1170055e-18 3.592317e-22 1.9724066e-19 2.64596e-15 1.0266845e-19 8.3273306e-19 9.332432e-19 5.4799966e-20 1.0 2.4942619e-17 3.997431e-20 7.1896484e-20 5.40928e-19 9.08474e-19 3.5996376e-17 3.012498e-19 5.812796e-18 6.250199e-18 2.4200303e-16 5.27634e-19 4.7054295e-16 4.5347163e-19 2.9706936e-21 1.5514553e-23 2.5307467e-19 3.394482e-18 2.0687277e-20 5.049286e-16 9.946722e-19 7.2468814e-16 1.1228118e-17 1.0656346e-18 1.846729e-20 3.6484479e-19 1.570747e-19 4.9070285e-21 7.019378e-20 3.344645e-20 1.4406832e-20 3.878188e-18 1.3400302e-18 4.252371e-20 1.16517695e-17 2.728512e-17 2.1996653e-17 8.572129e-18 7.997199e-20 1.7302772e-17 9.522739e-20 1.9237213e-17 2.7238948e-17 2.0580004e-19 6.913064e-18 2.1528555e-17 6.972316e-16 3.1694378e-20 2.3269804e-20 4.7976243e-19 5.050834e-21 3.752041e-15 2.022488e-19 1.5508574e-16 1.7967619e-17 1.1513699e-16 6.3655895e-14 5.1766183e-18 2.8784866e-18 6.0197235e-20 2.3571533e-18 3.341688e-19 4.0532072e-17 2.4284399e-15 8.316148e-21 1.5886208e-18 6.858201e-16 1.0501665e-18 1.4950621e-16 1.0 1.7692961e-18 3.2349377e-18 1.3199213e-16 1.1522508e-16 3.8884284e-16 4.352319e-19 1.5006747e-14 1.2441445e-16 3.3752675e-17 1.9124013e-18 7.1492683e-16 4.4548023e-18 1.0327307e-17 6.900515e-20 6.4500013e-19 5.1046927e-17 7.927949e-19 8.7551237e-16 1.9396212e-19 5.081672e-16 7.4050687e-19 4.5164834e-20 4.3259574e-16 1.3700568e-18 1.9926282e-20 1.5909048e-21 4.7292274e-20 1.9946542e-17 5.3814944e-19 1.1345412e-15 9.912141e-18 9.263904e-19 9.336963e-17 4.094379e-14 9.47686e-18 7.1449606e-16 3.9053091e-19 1.6536902e-16 1.4351116e-18 2.2470345e-17 1.092093e-16 9.5518074e-20 2.0476345e-17 5.2258442e-18 9.670634e-20 4.555281e-18 2.9896508e-17 8.441677e-18 1.8868539e-22 3.7529105e-17 1.833751e-18 6.9373942e-18 4.630495e-18 1.7626815e-17 3.7777204e-15 1.5687913e-16 1.3737567e-18 1.206084e-21 3.06873e-18 4.152136e-21 2.4680245e-18 2.7012982e-15 1.5564756e-24 1.2135438e-18 9.935663e-16 3.747317e-17 2.3087508e-18 3.464051e-16 1.0 2.0892906e-14 1.687248e-21 1.2202129e-16 2.0901892e-16 6.490288e-17 1.9287337e-16 1.7022055e-16 2.7387543e-14 6.5690264e-16 1.3764184e-15 1.0684518e-17 9.995795e-18 3.0493545e-21 4.0384175e-19 1.2530616e-15 4.9012757e-17 1.3789851e-12 1.6603881e-19 2.3307261e-15 2.1408824e-17 1.3885849e-15 4.0544245e-19 2.4142852e-18 9.628087e-19 4.591775e-20 1.195659e-19 6.111349e-20 3.677903e-20 5.149214e-14 7.5831094e-17 7.660004e-19 2.0193401e-14 1.4777915e-15 2.0189606e-14 3.7408347e-16 4.228286e-20 1.11947845e-20 5.82521e-16 1.2538343e-18 1.2838044e-16 1.7795008e-21 1.3722319e-19 3.3016585e-20 1.6485875e-17 5.846347e-20 2.9914174e-18 5.9894295e-17 7.6912266e-24 4.8139563e-20 8.7831046e-20 9.200618e-19 6.017683e-19 1.2216156e-19 3.0934107e-19 2.061234e-15 1.0083371e-19 4.456969e-22 1.5310573e-19 9.922141e-16 4.3929805e-21 1.0437888e-17 7.153162e-18 9.7729743e-23 3.6591226e-20 5.4631473e-22 4.1667856e-19 2.5516187e-18 9.962866e-18 1.0 3.551818e-23 3.9144235e-22 1.0362621e-18 2.583537e-19 7.4179324e-19 4.9476166e-18 3.972945e-16 2.0668524e-18 2.7254384e-13 1.2156217e-22 3.122192e-21 5.2552005e-22 5.5970986e-22 1.4277735e-17 3.5512238e-22 5.3228588e-11 3.670553e-23 8.513864e-17 6.7611895e-20 1.1228755e-18 1.5394202e-19 4.0820907e-21 1.1002496e-19 2.895959e-15 9.877247e-22 1.016127e-16 9.50884e-19 2.2279579e-17 2.3049318e-17 3.5585365e-20 1.2166661e-21 1.5259728e-19 3.5032554e-13 2.5378826e-18 1.1070557e-22 3.5271535e-21 9.604375e-22 3.758486e-23 5.2656602e-20 4.6189303e-22 1.8852308e-19 1.4027554e-18 4.5193788e-20 2.1991104e-20 2.2948188e-23 3.4075936e-21 1.5843191e-24 4.557121e-19 6.178908e-20 8.916188e-18 6.63556e-18 2.5759707e-18 6.261183e-16 2.3658983e-20 3.0349777e-20 9.756185e-21 1.4151372e-20 2.206202e-20 8.605037e-22 2.6893685e-17 1.8913762e-18 2.4703472e-21 2.0690229e-17 1.1713603e-21 3.4835806e-18 5.36224e-16 2.2562535e-22 3.7532732e-22 1.0 1.8813404e-19 8.46199e-18 4.1699455e-21 8.234204e-20 1.6561986e-21 1.8942258e-19 6.477555e-23 6.287283e-20 5.106421e-19 3.0573215e-21 3.4798164e-21 2.4591102e-22 2.6114909e-18 1.5505459e-21 5.333991e-21 8.3210985e-21 1.3945091e-15 2.283515e-19 4.801132e-22 1.32368726e-17 8.06001e-21 1.9846748e-23 4.7265944e-20 2.9692366e-18 9.305756e-22 1.5495135e-18 1.9789721e-18 4.1878342e-19 4.2196607e-22 2.8998111e-19 2.962019e-18 6.484307e-20 5.3414776e-16 3.0470483e-16 9.331911e-16 3.3314402e-20 7.983617e-17 2.8868785e-17 4.3105287e-13 5.1987735e-14 4.8077836e-15 1.745317e-19 2.610577e-16 1.0710297e-19 8.973483e-17 3.869469e-18 1.344029e-18 1.9012875e-17 8.660416e-17 1.6524227e-16 3.730945e-08 5.2321578e-14 1.3178033e-16 6.438362e-15 9.523215e-19 7.3146494e-16 4.3566165e-21 7.444563e-16 8.058998e-13 1.5358481e-19 3.219261e-14 8.58379e-15 5.1059617e-16 6.145306e-16 2.1082159e-15 4.0692108e-17 1.4444801e-20 1.9966792e-17 0.999998 3.2160402e-17 8.2160624e-14 1.2972063e-13 7.0694515e-17 1.1684672e-14 6.581493e-16 1.2560539e-13 2.1063105e-15 7.29344e-10 4.052876e-20 3.0015566e-18 2.3803873e-16 1.5989328e-17 7.537323e-15 5.642693e-16 2.8653024e-14 1.9694615e-17 6.915596e-18 5.4922395e-20 4.9212153e-15 6.953481e-17 6.508774e-20 2.9570627e-18 3.4704342e-19 1.9542894e-19 5.554581e-14 1.1300222e-16 5.4921405e-18 1.9738965e-12 1.6601723e-12 1.32519855e-14 1.6287984e-15 2.3727649e-22 3.6718467e-20 2.0240547e-19 1.10818065e-20 1.7839154e-17 3.690622e-20 2.6252433e-19 1.5210815e-16 7.240188e-18 8.304581e-20 7.605345e-21 6.522494e-20 1.0475271e-23 1.166663e-16 8.759086e-19 2.785492e-17 1.766093e-18 2.0486795e-19 2.3478843e-17 9.7968034e-20 3.609615e-20 1.4076364e-21 1.8441053e-18 1.1967822e-18 4.426656e-22 9.480071e-14 1.6910845e-18 7.314177e-23 3.299273e-16 7.4162275e-21 1.941659e-17 7.072418e-17 1.4975085e-19 4.7631226e-19 4.2572748e-19 2.747838e-19 1.0 3.149445e-19 4.8276246e-18 1.8950675e-17 5.5289435e-18 5.420426e-22 3.3705978e-18 1.8228075e-19 6.1759923e-18 8.3780164e-21 1.0391503e-21 1.4847677e-18 4.456391e-22 7.1638623e-19 1.0351029e-21 8.617654e-14 8.645884e-19 3.680596e-21 3.809076e-16 1.5444958e-20 2.9265775e-22 2.122123e-18 1.913234e-17 1.6425955e-21 5.409643e-22 4.1209956e-17 1.2974506e-17 1.5677823e-21 3.5251664e-19 5.456202e-18 6.4286025e-18 2.8711184e-17 8.0568027e-16 1.5799936e-14 5.839807e-23 2.0652872e-15 1.4073663e-16 2.9071256e-16 7.433666e-16 6.28475e-17 2.873065e-20 1.5935067e-18 7.004238e-20 2.734569e-15 1.3882486e-19 2.0647541e-20 3.217411e-19 1.5347579e-20 6.360167e-17 2.9321308e-15 1.9109229e-16 6.316091e-17 8.192031e-13 2.856918e-23 3.8270904e-17 7.7731993e-23 2.917583e-14 5.6942983e-12 7.541379e-22 1.1036916e-13 2.86626e-17 6.126487e-16 3.2151307e-18 4.86053e-20 4.829927e-18 2.217687e-19 1.8157893e-21 5.4866122e-14 3.5560554e-19 1.0 6.017831e-16 2.496405e-16 2.1465312e-17 1.4153189e-15 1.9682206e-11 7.026059e-17 1.903532e-15 1.0027457e-24 4.7339925e-20 1.937035e-17 1.1394498e-17 2.6218486e-13 1.434161e-16 9.370107e-18 1.5438721e-19 7.9261695e-18 6.3455892e-21 3.0945502e-15 7.835512e-15 5.0661945e-21 3.479418e-21 2.816096e-21 4.9925446e-20 2.6834684e-15 1.4757879e-19 3.7070903e-19 6.734382e-16 6.141794e-15 5.5579663e-14 8.3840115e-17 1.2620048e-20 6.5724003e-19 8.043363e-21 4.0033867e-18 7.3131875e-19 7.242452e-18 4.911193e-18 4.1413834e-19 3.8885095e-20 8.577129e-19 7.703134e-20 2.4390942e-18 1.2518352e-20 6.557803e-18 6.678648e-20 4.5945826e-18 9.112416e-20 2.39369e-17 4.8085566e-14 1.9159867e-18 1.2549688e-17 3.256006e-20 2.267287e-16 8.429685e-23 1.4621222e-18 1.31798056e-14 4.355154e-21 4.6030232e-17 2.268366e-18 7.722295e-18 1.3513247e-17 7.1966603e-16 2.1953493e-18 1.6542556e-19 1.8546149e-20 5.1047044e-16 1.8139575e-18 3.6761832e-17 1.0 3.1139838e-14 3.5392143e-13 9.220403e-18 3.55156e-16 2.3815839e-18 8.33528e-18 1.770113e-18 3.653394e-18 2.0506116e-18 9.246005e-19 4.181651e-15 2.0650152e-19 6.0842625e-16 1.1895667e-20 4.559714e-18 5.1742248e-20 1.4334976e-18 5.985857e-20 1.3875587e-21 6.937712e-18 8.125124e-19 8.91291e-21 2.7942323e-15 1.3803067e-18 7.256248e-19 7.750998e-14 2.8971495e-17 6.1139003e-15 7.504665e-17 3.0408184e-20 3.0503133e-18 8.0133535e-20 3.2607062e-19 9.609928e-18 2.08728e-17 1.2161818e-16 8.927108e-19 1.531886e-20 2.5102705e-16 1.4696474e-20 9.6029713e-17 5.6434565e-20 3.236484e-16 4.9969107e-19 1.832723e-18 4.637174e-19 2.5870297e-18 7.690965e-16 6.9104504e-19 6.3037906e-17 2.598012e-22 6.8923406e-18 5.349376e-21 1.4899104e-15 4.6331944e-12 9.353531e-19 1.0809474e-17 6.6794226e-17 2.6655186e-18 4.838987e-16 1.0539692e-14 6.9597964e-17 5.7459976e-17 3.486076e-20 2.8766179e-16 1.0108546e-14 5.313432e-16 6.1914823e-13 1.0 2.8593957e-12 2.420457e-16 3.247121e-15 2.4281503e-17 2.4190263e-16 3.701394e-20 2.2500696e-18 8.4324764e-17 3.2575003e-18 3.4940635e-13 1.1672762e-18 4.096705e-15 1.3158585e-17 4.6050758e-18 1.6783332e-19 2.732375e-18 4.7768994e-18 7.5809523e-22 5.4572554e-21 7.2574623e-20 1.2459135e-20 1.2357412e-13 1.515185e-17 6.2693477e-19 1.7608984e-12 7.8355557e-19 4.2632155e-13 1.2966918e-13 1.1530522e-18 5.8426575e-16 1.2928853e-18 2.248956e-16 3.5654856e-17 1.6100664e-17 5.1954946e-17 6.241598e-18 1.735266e-19 2.388198e-18 5.9907225e-20 7.3218165e-18 2.6215092e-20 2.1932275e-14 1.06338535e-17 2.1162632e-17 1.3178774e-18 6.485929e-18 2.0511165e-14 1.5031195e-16 5.4180854e-16 1.0072021e-21 4.645928e-16 2.9833493e-18 3.1113313e-17 9.39218e-12 4.4927846e-21 3.06419e-19 7.4576875e-18 9.796738e-18 1.8607059e-16 3.4670107e-17 1.619009e-17 1.871263e-15 8.0476634e-20 6.050086e-17 1.2332952e-16 1.2197501e-16 2.9169963e-12 1.428311e-12 0.9999763 6.654905e-17 9.587799e-13 3.9379576e-16 1.4729041e-18 1.9912613e-19 1.903223e-19 7.414065e-17 6.6468525e-20 2.764792e-12 6.142511e-18 1.22810875e-14 3.1293494e-19 3.938367e-14 8.027466e-18 4.169284e-18 1.9349727e-19 3.1328883e-17 2.7609262e-18 8.232611e-18 5.8590986e-19 1.3666556e-15 7.312913e-17 1.6957494e-17 4.1639814e-15 5.999284e-17 0.0002604324 1.477351e-16 5.0044914e-16 3.1138248e-17 3.6812717e-20 2.1972734e-13 4.133244e-17 6.7798444e-18 5.8829593e-16 2.8342736e-16 1.7466929e-22 3.937884e-15 9.6815545e-21 2.0753942e-17 1.444222e-19 1.6861695e-18 2.5068914e-17 1.532549e-17 1.01893e-15 5.4762017e-17 8.2945346e-15 6.8990397e-13 8.326535e-12 3.0007535e-19 1.2877149e-14 1.1630753e-19 1.5648114e-15 4.791752e-12 5.833471e-20 6.425404e-16 3.579304e-17 1.7987506e-17 2.9583987e-15 7.91892e-17 3.4657395e-18 5.9853466e-15 1.0405034e-20 4.046915e-15 2.518178e-20 1.801184e-14 6.689805e-16 1.1795883e-14 5.9874686e-14 1.0 2.271407e-09 3.728225e-17 9.960396e-18 7.217375e-20 6.5026757e-18 8.6571516e-16 1.6537398e-17 6.1266373e-13 8.424503e-17 7.4236976e-14 3.4261726e-18 1.0192716e-16 1.4091462e-21 1.8174173e-14 1.1979526e-16 1.6874686e-19 1.5611927e-22 1.3229401e-17 2.0479157e-17 4.533826e-15 2.0292668e-17 1.0574159e-15 2.292326e-15 2.2617926e-16 5.887772e-11 6.3940607e-15 9.023777e-16 1.1615967e-15 8.004349e-18 1.1042168e-15 2.3473685e-14 6.890662e-17 1.00101574e-14 5.266937e-16 3.4554161e-18 1.1765358e-16 7.3182355e-20 1.0781106e-16 1.1469002e-21 2.8333394e-19 1.8216494e-16 1.7236366e-17 9.610158e-16 1.4362414e-16 2.6945262e-15 2.8774635e-12 1.9550176e-14 7.57503e-21 1.4574474e-14 4.992678e-20 1.3930729e-16 7.419823e-13 4.8813014e-21 4.1018435e-18 3.0855678e-15 1.5470926e-17 2.3175251e-15 2.797237e-17 4.966453e-17 2.2271297e-13 7.532269e-21 8.127813e-17 5.8637076e-18 1.5067444e-13 7.0906524e-17 1.5336776e-17 6.2753665e-13 4.710141e-13 1.0 4.2992057e-17 2.885655e-18 1.1543444e-21 9.376609e-19 1.4122987e-15 6.2156016e-19 3.3459367e-11 1.1260197e-18 2.0749144e-14 1.5825663e-18 1.5580249e-18 3.1930568e-18 9.101874e-16 6.435026e-14 5.019837e-19 1.8321859e-17 1.0860809e-17 1.6403633e-17 2.5430992e-16 1.6331149e-16 6.391372e-17 1.991461e-17 1.35812255e-14 1.9963011e-12 1.0972241e-18 4.5346646e-19 7.135012e-18 1.8437569e-22 2.2040075e-19 2.3859645e-20 5.734977e-19 3.139841e-16 5.0221788e-17 7.922858e-22 2.6593988e-22 1.1399693e-21 1.2138353e-18 1.2908036e-21 5.2833948e-17 3.8752938e-16 4.2869346e-19 1.09092505e-20 1.4292768e-18 2.3661943e-14 2.1866712e-20 5.159433e-16 6.6808037e-23 1.4827972e-19 4.6023166e-19 4.2800598e-17 2.3678208e-13 9.23519e-25 3.1067738e-19 2.4787496e-20 2.8627946e-18 1.48921e-17 2.0915988e-18 2.2496123e-20 7.135841e-22 3.785533e-18 9.049293e-18 2.039105e-19 2.3749262e-17 1.30264345e-17 2.0854485e-18 4.1585332e-15 5.1386073e-18 5.782679e-16 1.0 9.500129e-21 6.7246284e-23 1.861689e-20 1.7818588e-20 6.182232e-20 4.4918515e-17 5.654908e-17 5.631723e-17 1.2792492e-22 8.836227e-22 1.4195714e-19 8.0428267e-17 2.4380047e-19 3.5530513e-23 7.134892e-21 2.1957298e-22 6.9227125e-21 4.0319566e-17 2.5266751e-19 1.5218404e-18 2.4493169e-15 8.937539e-18 3.6682144e-14 4.1723767e-15 1.0275447e-15 8.178008e-14 5.6443736e-23 3.385094e-18 1.2001738e-19 6.873172e-11 8.483232e-16 8.901142e-16 1.8064417e-21 9.78938e-18 3.731219e-21 6.0635456e-17 1.2087584e-16 6.848494e-18 7.709631e-20 9.7934735e-21 5.1419433e-17 4.1085354e-08 1.7281152e-15 9.23488e-16 4.1686383e-15 5.3314505e-20 2.4251419e-17 1.1173532e-21 3.239966e-17 1.3004467e-13 1.4472648e-21 4.774543e-13 2.7221288e-17 2.6683916e-15 1.3840771e-18 2.2968323e-16 3.7101052e-18 1.5781577e-20 1.0602284e-20 2.0411767e-06 2.2871364e-16 1.3114407e-14 6.9702755e-14 2.4700247e-15 2.8754474e-16 6.5461347e-17 1.3113e-15 3.8631588e-17 1.0 1.403374e-21 2.771172e-18 2.666374e-17 1.0260411e-14 7.461227e-15 9.071198e-16 9.623171e-18 1.5174695e-18 1.2825982e-19 1.681928e-20 3.1420385e-13 4.981255e-16 6.413685e-20 5.688581e-22 1.7209732e-20 3.5873363e-20 1.1820792e-14 3.0141036e-21 8.0291285e-16 1.474678e-14 9.624724e-16 1.5950894e-14 6.3813667e-22 7.0612286e-25 2.2107395e-23 2.671325e-22 8.543473e-24 2.395747e-23 5.1224454e-24 3.3013135e-23 1.2592709e-24 9.791446e-23 7.938386e-20 8.976537e-23 1.5542972e-25 2.0691239e-26 2.3072979e-18 3.4057597e-22 2.9859118e-10 7.983296e-21 2.221536e-22 3.974062e-20 4.6596335e-22 1.7890961e-23 6.221434e-20 2.8822605e-22 1.3871301e-21 1.0945864e-25 6.0312138e-21 6.1986004e-19 3.1519608e-24 3.8012259e-22 2.6648279e-25 1.070836e-22 1.20355215e-20 1.2408034e-23 5.3045027e-23 2.6360557e-23 2.9646763e-21 8.966958e-22 3.942182e-23 3.7770065e-19 1.826231e-20 9.3514796e-21 5.364249e-22 1.85156e-21 1.1514882e-24 4.6177825e-23 1.0 2.7741037e-24 1.2414722e-21 4.3580893e-24 2.2444595e-21 6.8919973e-26 7.959738e-19 3.4775702e-23 5.544331e-23 6.480277e-22 1.005895e-24 1.590938e-25 9.095759e-25 5.6338785e-22 7.8626675e-23 1.8171396e-22 6.840195e-17 2.1187448e-22 1.8886245e-23 1.0420396e-22 5.4724495e-22 4.452801e-22 1.7108014e-20 2.4549842e-16 1.5423036e-15 7.8056474e-23 3.9610346e-18 1.2981924e-19 5.1551757e-19 1.5221953e-17 2.267486e-16 1.9287572e-22 7.029285e-22 2.9443305e-19 3.8466848e-19 1.181069e-15 4.0416015e-22 5.0036935e-22 2.1527218e-19 5.0502804e-19 1.9039702e-20 2.5653044e-16 6.39464e-17 2.6633955e-17 1.0313533e-16 3.441225e-17 7.366187e-27 4.401204e-19 3.4771298e-17 4.425856e-25 1.4935117e-16 6.13886e-17 3.6648577e-16 4.2881123e-19 7.928256e-18 1.7111743e-19 6.552742e-21 2.193956e-21 8.743349e-18 1.835043e-21 3.284339e-18 1.2015225e-16 1.1190361e-17 2.9830168e-18 4.800833e-17 1.4128253e-17 2.4166326e-20 1.7775088e-18 1.9706493e-24 1.0 5.495155e-19 1.0786095e-14 1.9128346e-15 1.7125147e-17 6.041902e-20 3.234564e-20 2.1034146e-20 1.5539566e-21 3.720216e-15 3.6812293e-20 9.832604e-26 1.0762454e-23 3.3650428e-21 1.6548959e-23 2.4882107e-15 7.5081075e-22 6.083543e-16 8.515441e-13 5.6716233e-19 7.580353e-18 2.2756488e-16 1.2479166e-19 1.2534194e-16 4.3429836e-18 4.2634877e-17 3.9769213e-16 1.0646593e-18 2.4650053e-17 7.697593e-17 6.854584e-18 1.4823518e-17 3.0302937e-19 5.4522457e-19 8.817098e-23 2.1231503e-19 1.5548306e-18 6.936045e-18 3.1357097e-17 2.4518236e-14 1.4421061e-14 2.4471515e-14 8.6155943e-19 1.4965555e-18 4.4178446e-17 5.0157573e-20 3.9139636e-17 1.0943665e-15 8.244352e-21 5.4303477e-19 1.5567922e-14 1.4275978e-19 9.1482e-17 7.711319e-17 4.3369596e-17 1.9044815e-17 3.7169599e-19 1.1004667e-17 1.222973e-18 1.4157703e-18 1.7035287e-18 4.241553e-18 3.4445987e-16 4.650341e-17 6.6537695e-15 5.2744658e-20 1.901671e-18 1.1803042e-21 5.493161e-20 1.0 1.0811537e-17 3.3046626e-17 1.7189655e-20 1.4901927e-10 1.0719082e-12 1.3245447e-19 5.5201207e-17 3.7787517e-18 1.3601106e-18 7.099229e-18 7.5928015e-19 4.9026545e-19 3.764336e-18 1.3144996e-16 1.540033e-16 2.8771567e-19 8.0286997e-16 4.513115e-17 4.6355116e-17 5.7445923e-18 1.729849e-16 2.4651488e-15 4.9429606e-24 1.7633389e-19 2.0975974e-20 1.3816987e-17 2.2075439e-17 9.812415e-17 3.0299542e-22 6.702691e-20 1.7615448e-18 5.8569145e-17 8.0144583e-19 8.1564815e-21 7.067037e-21 2.9134004e-22 1.0024166e-17 1.492115e-18 1.9742455e-17 6.177274e-15 1.87352e-15 3.8411667e-14 9.478393e-17 4.1119513e-24 5.3139336e-18 6.275246e-15 3.2266582e-24 2.5127163e-14 5.8311654e-17 1.8342297e-15 7.16578e-18 4.596654e-17 3.6153632e-19 1.802855e-21 9.675089e-22 4.3857956e-18 9.305471e-22 1.3271045e-18 1.9675699e-16 1.624757e-17 7.595141e-19 1.7539896e-16 2.5639936e-16 1.0175063e-19 4.6286053e-18 9.919263e-24 1.1139865e-14 1.03806336e-16 1.0 2.0804614e-15 3.5426428e-16 3.3674494e-18 5.948271e-14 2.6693736e-19 7.9958146e-23 1.1045045e-13 4.3451027e-19 1.646975e-21 8.2360173e-22 3.3883946e-22 2.1893933e-19 1.00831285e-14 2.6597847e-21 2.482283e-17 7.261389e-14 1.7947127e-18 7.051532e-17 3.9445918e-14 7.86699e-15 8.143728e-15 5.290958e-17 2.1137468e-16 7.794067e-17 1.971566e-17 1.0217174e-14 7.9668584e-16 1.5909262e-17 1.06973537e-19 3.687299e-21 1.9662634e-17 3.3715975e-20 9.734081e-18 1.464299e-18 2.6022797e-20 1.5832614e-17 4.3107645e-18 3.652936e-17 2.3174885e-16 6.3156904e-14 5.505827e-22 1.1586746e-17 2.3544535e-21 4.60208e-15 3.4815052e-12 1.1601919e-24 1.6069064e-17 1.1665551e-18 1.0921436e-15 1.9835248e-15 7.250231e-15 2.255454e-14 9.695399e-09 1.2525744e-22 5.3593866e-16 4.151968e-18 4.8560412e-14 6.1284507e-16 1.5775787e-14 4.9277366e-12 4.9772685e-15 2.3523664e-10 9.462276e-18 5.726785e-17 1.3977855e-21 1.1721857e-16 7.941124e-17 1.2203313e-17 1.0 2.5482337e-18 6.4907303e-18 4.046112e-19 2.6636814e-16 1.1231276e-19 3.9033345e-16 1.5921809e-12 1.2305393e-19 2.5792216e-21 1.6757503e-17 2.1028241e-17 4.448732e-14 8.309982e-16 5.960758e-16 5.35672e-16 3.9541374e-17 1.0700204e-11 2.7588283e-20 1.2212972e-14 3.3206006e-15 1.869701e-22 1.5063597e-20 2.5856518e-23 3.1503614e-17 2.835386e-17 2.2209582e-16 3.6625386e-23 3.9968758e-23 5.724016e-23 5.724276e-16 1.0422285e-18 1.6897479e-20 1.6627768e-20 1.9090093e-21 3.101738e-21 7.109167e-20 1.0920597e-16 4.5286287e-20 3.327513e-14 3.0704672e-23 8.25578e-20 1.7940972e-25 2.224333e-16 2.0648618e-15 2.893879e-25 2.3230903e-17 7.289263e-19 3.6124066e-15 2.3602047e-17 4.2951993e-20 2.175296e-22 4.0422934e-23 3.2762398e-20 1.3869245e-16 7.3588187e-22 8.416188e-16 6.256064e-19 9.284034e-19 2.1545615e-18 1.568976e-17 1.242488e-18 4.1384028e-17 2.7023455e-18 8.155127e-25 7.497332e-19 5.158492e-22 6.298383e-17 7.988217e-17 1.0 2.3798548e-21 3.7289407e-22 3.1989315e-19 5.2324106e-24 1.3562374e-15 3.0377507e-17 9.056851e-23 6.586058e-25 7.643843e-24 6.0025916e-21 1.7305263e-19 5.267808e-20 1.3311822e-16 3.8083617e-17 3.562642e-18 8.029917e-16 2.334311e-16 2.6114968e-21 1.3313834e-19 2.907942e-19 2.8373957e-19 1.2292739e-17 7.875111e-19 1.2859444e-19 9.760173e-20 2.246628e-20 1.5676622e-14 1.6267583e-19 4.6721715e-19 4.301932e-23 5.991486e-17 4.091139e-18 1.5138927e-15 1.6495429e-18 9.0833075e-18 7.8476226e-16 9.734601e-18 5.984875e-20 1.3950869e-18 1.8254249e-18 1.1235953e-18 3.7166605e-20 3.0231876e-15 1.928202e-18 7.4936925e-21 5.5092628e-18 1.089308e-19 4.3207435e-17 4.588715e-18 6.9454406e-19 3.4927466e-19 3.7274247e-19 2.8853833e-19 1.2554196e-16 1.1299692e-19 4.6515096e-18 9.698975e-19 4.2994645e-17 5.844773e-17 1.6849496e-16 7.9525706e-20 6.1349377e-21 2.750212e-20 1.417136e-20 4.450228e-13 4.871575e-20 4.7837535e-19 2.040052e-21 1.0 7.7555327e-19 2.4085797e-20 2.7698692e-20 3.2029105e-20 1.7255147e-21 8.695631e-19 8.017857e-18 8.643369e-21 7.416995e-20 2.7483076e-17 1.20308995e-17 4.352829e-21 1.7239456e-17 3.1501208e-17 2.8982283e-18 8.5906e-16 1.6087133e-20 2.5603326e-20 2.6249617e-20 1.408707e-19 1.0752689e-15 1.4564378e-19 7.267225e-17 3.4748094e-17 2.4701625e-18 3.6745168e-16 1.7751043e-17 6.173819e-20 1.645329e-22 1.459162e-20 3.1398036e-17 1.3575973e-17 2.8513752e-16 2.0441078e-16 5.584548e-18 3.6331606e-17 2.2257799e-20 2.470523e-16 1.408848e-17 8.624102e-21 6.5942994e-19 7.0188254e-17 1.3684047e-20 4.967926e-19 1.6023836e-16 8.870405e-22 1.2895133e-15 2.1548344e-18 2.6613107e-19 4.021553e-19 1.2332791e-19 8.698936e-18 1.290501e-17 1.89657e-19 2.153872e-19 1.2274611e-16 5.727982e-17 1.8837211e-19 1.4252664e-15 1.0532466e-20 8.196097e-20 9.313248e-22 3.4171367e-20 1.51138e-11 2.0878712e-15 9.143281e-17 7.1466557e-22 3.6476203e-15 1.0 3.1518246e-19 7.963144e-18 8.166172e-19 5.1664707e-21 2.1780323e-18 2.06237e-19 8.92223e-21 1.3189335e-18 1.6801704e-13 1.6601611e-17 3.9990612e-21 3.702259e-18 3.2298328e-18 5.0475814e-18 1.9177343e-18 9.899659e-21 3.5518098e-20 1.9995999e-23 7.3149727e-19 7.2802e-20 8.172806e-20 4.0384175e-19 9.842612e-21 5.818685e-22 1.5074017e-18 8.9559556e-21 3.0489722e-20 2.531649e-23 1.1489175e-16 3.3347623e-18 8.12469e-19 4.2852956e-21 6.859386e-21 1.5673891e-19 4.5860328e-20 6.4224506e-20 9.8135085e-23 7.88052e-19 6.481e-20 3.7007124e-22 6.7083834e-16 2.9965739e-22 1.2432128e-19 1.1825996e-19 2.071941e-21 3.834368e-17 4.5451243e-19 3.107915e-17 5.882914e-16 3.0956606e-22 1.5961856e-19 1.086702e-18 3.2259281e-19 1.0291559e-17 2.4434527e-18 3.812698e-13 7.393105e-18 1.8398972e-17 1.7060773e-19 8.623036e-24 1.5001943e-21 2.2260494e-22 3.0067711e-18 4.040619e-20 1.6494645e-14 1.3159368e-20 3.4323228e-17 2.377896e-20 1.0 2.1637108e-21 9.280311e-20 4.4034285e-23 5.028974e-21 2.1421476e-18 1.4545992e-20 3.1492495e-21 1.2062842e-17 5.5367043e-21 1.0652102e-21 1.3479159e-18 7.094669e-21 2.7839318e-14 1.2695871e-18 3.4816337e-22 2.7007128e-21 6.390905e-18 6.1370475e-20 2.8689589e-19 6.175626e-22 3.8317736e-20 2.3450087e-19 2.8927588e-14 1.2090372e-20 5.3492148e-20 2.9016798e-20 2.5430596e-24 9.067334e-21 6.576084e-20 3.4332113e-18 4.210231e-18 3.7049536e-20 1.9890864e-15 1.2892008e-19 7.7279965e-22 4.67124e-21 1.5708308e-19 7.610623e-22 4.2613347e-20 3.369272e-19 2.241636e-21 7.5210636e-23 2.7828368e-17 1.2037254e-22 6.366982e-19 6.564865e-17 2.8962045e-21 3.4838825e-20 2.8559225e-18 3.6404017e-21 6.1350003e-16 1.1248851e-20 4.343261e-20 3.8418145e-20 1.1568805e-17 6.260041e-22 4.498985e-17 2.2235744e-19 3.324141e-20 4.0643463e-21 1.5956577e-21 4.1926195e-17 1.3692963e-22 1.3032733e-19 7.4491986e-22 9.994727e-18 1.2308687e-18 6.7796756e-22 1.0 6.349637e-20 4.8694523e-22 1.384646e-20 6.8634995e-20 1.1803635e-20 1.1556013e-19 4.552293e-18 1.9017511e-17 1.2807683e-21 1.8165108e-20 1.5799716e-19 1.06991494e-17 7.6029785e-18 1.24017965e-14 2.915037e-13 2.0425183e-23 8.314829e-18 5.0316854e-19 2.8918287e-16 4.759227e-15 6.5881024e-15 1.183704e-21 2.4938624e-17 9.485779e-23 1.0204932e-16 3.5516494e-19 2.4534253e-20 5.4388e-21 3.2551633e-19 2.3845217e-17 1.5089738e-16 3.876218e-12 8.641116e-16 2.0599185e-10 3.8561025e-19 7.4236303e-17 3.1074298e-24 7.1956995e-16 9.494248e-16 1.8833375e-22 5.9452514e-15 1.5680748e-14 2.1163944e-15 1.1105649e-17 5.641678e-17 7.5297734e-20 8.7814295e-20 1.1868522e-19 7.2379745e-15 6.9379696e-20 6.0062513e-15 3.700355e-17 2.3992926e-18 4.920045e-16 2.5586124e-14 7.510752e-13 1.4597894e-15 2.4424044e-14 1.6980614e-23 8.68871e-17 5.040333e-18 6.963999e-15 1.2110259e-14 8.188895e-16 7.6929266e-17 2.457623e-19 1.9378608e-19 1.1610628e-19 1.0 2.629967e-17 1.2830233e-20 2.0612422e-22 1.01004205e-20 9.701703e-21 6.400064e-15 3.96558e-20 8.882516e-17 6.616635e-16 9.333264e-16 2.5289058e-15 6.980669e-17 3.7175545e-12 1.3297124e-13 3.538798e-21 7.4225505e-18 1.095092e-17 4.159076e-15 4.0404156e-14 3.5493523e-16 1.7442977e-20 4.161319e-20 4.415763e-21 4.066344e-15 8.4926335e-21 2.2237004e-20 1.3310394e-17 1.2754151e-21 1.7590341e-17 4.118153e-16 1.1624433e-15 2.7573824e-16 2.014785e-14 1.1911917e-21 3.6825957e-18 8.07502e-23 2.974391e-13 2.5914614e-13 7.4038635e-24 3.8493036e-16 1.9634162e-16 4.934715e-14 2.5418077e-17 1.1961803e-17 9.269707e-18 5.2676976e-17 1.2357031e-21 4.5110423e-16 2.0852632e-20 1.3289112e-13 1.22904414e-17 2.4845104e-16 1.318884e-15 3.3491863e-15 2.8472755e-10 7.593823e-17 5.321871e-16 2.4586763e-24 2.2445682e-16 1.3599141e-17 8.323778e-18 5.0980144e-11 7.7162154e-14 6.6126162e-18 3.049903e-20 4.198599e-20 2.9999508e-20 1.1813348e-15 1.0 1.6881687e-21 3.329293e-20 1.378932e-20 4.9598804e-18 6.174043e-16 2.634301e-19 1.3962756e-16 6.0091593e-16 8.675202e-16 7.1506624e-13 2.5653245e-08 9.262234e-21 3.1563348e-20 2.4869343e-19 5.0066914e-22 5.0538622e-20 1.7887236e-20 5.8903736e-22 1.0799571e-19 5.257434e-19 2.0969035e-18 2.752995e-19 5.949167e-18 4.170144e-25 4.642019e-21 5.5835043e-18 4.258938e-16 8.336712e-18 3.9001175e-17 5.584396e-19 3.0753665e-16 7.159644e-22 2.6121944e-21 3.3129646e-19 3.3203363e-14 3.4500127e-20 1.9917809e-16 2.444279e-13 6.3596933e-21 1.6669257e-18 2.2618193e-23 2.0444636e-19 3.1464157e-21 6.387163e-22 2.6186074e-15 1.5544732e-20 1.3722337e-20 3.7291923e-17 1.10579895e-20 1.0767731e-19 1.0216791e-20 4.2403248e-16 1.0030921e-20 1.2157453e-17 9.557333e-23 3.0786115e-20 4.8932278e-23 4.3295526e-25 3.239027e-17 7.3319196e-22 2.1569306e-19 1.8162031e-23 4.5295045e-15 6.8470537e-19 9.062765e-22 6.3338635e-20 1.4112341e-20 4.1985926e-23 1.0 1.0350724e-19 5.913801e-18 1.7786211e-18 3.9644187e-19 2.560277e-17 9.336828e-23 2.248315e-22 8.6540344e-19 2.4795856e-15 2.0445671e-17 1.1516125e-22 4.367849e-21 5.0946716e-18 5.6562286e-23 2.835474e-16 6.5369346e-22 7.2643425e-17 1.1934345e-20 3.0996105e-20 1.0296614e-19 1.019028e-21 7.1269967e-23 3.3573942e-25 1.6775788e-18 3.126043e-20 8.526447e-16 1.6225564e-19 4.2460796e-21 1.3246772e-17 3.9080345e-20 2.089571e-20 7.405477e-23 2.5804137e-20 1.8762342e-19 2.8141614e-22 2.2696767e-15 2.5849063e-20 1.207727e-22 1.489735e-14 1.5411454e-22 3.4664516e-19 3.632824e-19 2.7850612e-21 6.6934156e-20 6.876419e-20 2.053871e-19 1.0244058e-15 6.322053e-22 7.790912e-18 7.1898605e-22 1.0299253e-16 3.9883466e-23 2.3622406e-17 2.5817638e-19 6.866998e-23 1.3258929e-21 4.6611582e-24 7.830331e-18 2.9558838e-22 1.8564916e-19 4.686797e-24 4.1551137e-14 1.4890209e-19 8.948105e-20 2.5063999e-20 1.7925607e-22 1.5796913e-23 1.9399393e-19 1.0 4.261219e-21 1.0344989e-21 2.3983788e-17 8.79914e-18 6.3221705e-23 3.7990535e-21 2.8772035e-17 4.7082004e-19 3.07714e-15 1.4057544e-19 2.1175126e-19 8.476467e-17 1.35862216e-11 5.733252e-18 1.0640665e-18 1.7810179e-18 3.6223922e-18 1.3518698e-19 3.573027e-18 2.0101649e-20 1.2901461e-18 1.8704275e-23 3.5486966e-19 1.820815e-17 4.232266e-15 8.912661e-16 5.009307e-15 5.0678478e-17 1.4880206e-13 9.1163195e-18 1.134798e-20 1.051291e-14 3.1706385e-18 2.0657725e-18 2.348264e-14 1.0268479e-21 1.0559701e-20 1.1991219e-18 4.4679e-21 9.478244e-18 5.7953544e-16 5.0426352e-20 1.7310342e-12 5.8270177e-21 6.7883126e-20 1.0583753e-17 3.8906608e-20 6.4398576e-16 6.799136e-19 3.2473039e-15 1.7051412e-18 1.6813299e-14 1.3136799e-20 5.492171e-22 2.2108019e-20 1.8963169e-19 1.8380092e-18 1.1440163e-21 2.677456e-15 1.4812402e-22 5.1836585e-16 2.5091933e-20 5.965227e-20 5.0058613e-18 6.192192e-20 3.1762886e-20 1.6348863e-15 2.54792e-20 1.0 4.4038116e-17 2.7590447e-16 4.438536e-17 1.4034352e-18 1.5322008e-21 2.0709032e-16 6.756924e-14 7.104891e-18 6.68423e-20 3.5867886e-20 7.7684984e-20 2.899947e-17 1.5161965e-18 1.6595776e-19 2.7071328e-18 1.5071403e-14 8.707344e-20 1.4188244e-19 5.71412e-20 5.7941303e-19 1.0251288e-23 7.184192e-20 5.57976e-17 1.4776941e-17 8.820955e-18 1.0695319e-18 2.0387636e-18 4.634327e-19 7.647887e-19 4.468226e-20 6.0275104e-21 5.841362e-17 4.5781393e-17 4.3174295e-18 1.241981e-18 2.0921792e-20 8.349146e-20 1.04047955e-20 1.7351022e-17 5.219807e-18 5.782355e-21 1.0865788e-16 2.940226e-17 5.386862e-20 7.562874e-21 2.3442114e-20 1.7771148e-19 4.04449e-20 7.3604566e-18 4.2191037e-19 1.4722322e-15 4.436722e-19 1.2860656e-21 1.373456e-21 2.590056e-22 2.0117789e-17 7.6808373e-19 1.4081817e-17 1.5040336e-21 2.460835e-16 6.68584e-19 4.8415047e-21 1.8788206e-18 1.9730603e-20 4.5005747e-20 5.988969e-18 5.320436e-21 3.202106e-19 1.0 1.3448805e-18 2.1314267e-19 5.3211467e-21 1.1639891e-20 2.131152e-18 1.7688913e-16 2.9076928e-15 4.3804683e-19 3.4223452e-18 4.7446455e-18 2.5439878e-19 3.5006734e-17 2.1742146e-17 1.2809566e-17 8.342473e-17 3.2961733e-19 1.797558e-16 1.620003e-18 7.2591e-19 2.85824e-21 1.6985341e-16 1.7550136e-16 1.5934811e-12 1.0036454e-16 3.9492965e-18 5.962402e-14 1.2253933e-16 1.948172e-18 6.015079e-15 2.1786824e-16 5.0645542e-20 4.4660834e-18 2.5932698e-14 1.5513055e-18 4.6979988e-18 1.5841684e-16 1.5048106e-18 1.5141915e-17 5.676654e-16 4.652113e-18 4.2220663e-19 6.915935e-19 3.4406812e-16 2.9812817e-16 3.2128135e-18 7.0801722e-15 6.7073596e-16 2.4451035e-16 5.747971e-17 5.661386e-16 4.7543566e-18 1.0658749e-17 6.91468e-16 1.4965792e-17 7.963725e-17 1.7310563e-17 2.3533383e-15 8.148054e-21 2.2669089e-14 1.002506e-14 5.5928173e-19 7.614588e-18 7.0833494e-18 1.9539911e-19 3.538327e-20 1.5722892e-17 1.2848019e-19 1.8591492e-19 1.0 1.0562584e-15 1.0790111e-17 3.0409984e-16 2.1137388e-16 1.5335013e-16 4.8577546e-18 2.5255555e-21 1.832994e-20 2.3660287e-09 2.0764681e-20 2.585402e-18 2.9501516e-21 1.7561856e-16 1.93648e-18 2.2655307e-17 8.7776875e-18 7.861596e-21 1.3660168e-19 1.0805153e-23 7.988079e-21 6.7807976e-19 1.6524289e-16 8.19888e-20 4.008043e-20 2.5120704e-17 2.7683812e-19 2.9108596e-20 1.4556802e-22 1.7716592e-19 1.1439116e-21 1.1122917e-15 1.5201426e-15 5.6500237e-21 6.8996996e-20 6.507682e-20 6.6240175e-21 1.0014e-17 3.9620318e-18 7.86883e-20 1.0583712e-17 2.946262e-20 1.2169207e-18 1.0826228e-17 2.8219583e-20 1.0065122e-18 1.388185e-19 3.670743e-17 2.1325406e-20 6.465947e-17 1.6325025e-19 1.7179091e-21 3.9051564e-21 1.7412909e-22 2.137398e-17 4.444558e-22 4.032497e-16 6.705376e-20 9.125137e-15 1.3079906e-18 1.6399714e-22 1.7162415e-18 1.3646417e-19 1.4208479e-22 1.0397387e-18 1.5782075e-18 7.896314e-20 2.2470909e-20 5.472905e-16 0.99999964 1.9443029e-19 3.2115604e-20 2.5006355e-17 6.442866e-18 5.03933e-19 5.1632155e-15 3.853144e-13 1.4401667e-20 6.5414454e-16 2.6863321e-18 4.2180322e-16 9.091538e-15 2.0390714e-14 2.5472262e-23 1.1704637e-19 6.467734e-21 1.9066778e-16 8.264738e-18 4.7199818e-20 8.199251e-21 7.43335e-17 3.0922576e-17 4.497333e-19 7.2391285e-13 2.1151478e-11 7.4680076e-14 1.1862264e-21 5.314167e-14 5.707423e-23 6.6616935e-15 5.462854e-15 3.2528496e-22 1.390547e-14 3.8643675e-17 2.708406e-16 2.1045646e-18 6.767496e-17 5.41992e-18 1.0141355e-18 3.55616e-21 1.8517023e-16 6.822903e-21 1.875551e-15 1.9760245e-16 3.0335948e-16 2.5281445e-15 1.4097758e-13 9.410173e-14 5.889785e-16 2.2632082e-16 8.3397756e-23 9.556715e-15 2.2453696e-19 2.0057417e-16 2.6664187e-14 1.1555646e-14 2.886215e-19 8.4174463e-22 2.9733027e-20 9.700441e-22 2.0678752e-13 1.30355815e-17 8.801543e-22 9.023331e-23 3.6298753e-18 9.051585e-19 1.6818303e-14 1.3496241e-17 1.0 7.977476e-15 3.897814e-16 3.167293e-14 2.2145095e-18 5.3256373e-20 6.161479e-16 9.496569e-23 2.4479961e-18 5.811953e-18 7.961021e-17 2.635301e-16 3.281725e-16 1.26821595e-20 5.826046e-19 6.9802056e-20 5.662204e-17 1.9471679e-19 1.07236416e-16 6.8235344e-19 3.0482877e-19 5.4112196e-19 1.1656889e-16 9.870066e-12 8.739565e-18 1.2336104e-16 5.3740942e-17 2.0735989e-16 1.003359e-22 1.2923893e-15 2.6166032e-13 4.0697104e-23 2.2988837e-16 1.3406507e-16 1.17791e-16 6.801945e-17 7.3205675e-16 4.9638227e-16 3.0291584e-21 8.1452656e-19 3.258246e-14 4.9890472e-18 1.3558384e-16 2.9460939e-12 2.3327735e-14 1.2336249e-14 2.1876259e-17 5.8253467e-15 7.9000667e-16 1.1158391e-17 1.0416108e-21 3.8242775e-15 1.591298e-15 3.3501828e-15 9.006036e-15 8.239024e-17 1.728492e-14 1.2246997e-19 3.2899192e-18 2.7907203e-20 8.369791e-17 1.6995166e-19 1.2494343e-22 3.0523007e-20 8.309457e-21 9.351024e-21 5.2487472e-15 5.1446214e-20 9.83981e-18 1.0 5.009777e-18 6.32996e-15 4.9320022e-15 2.1305507e-18 6.1058595e-17 5.1171044e-20 2.253858e-18 1.1576459e-16 1.1035961e-18 7.2751296e-17 1.3374735e-16 7.53343e-18 5.9027575e-18 6.0278125e-20 6.791774e-19 4.126524e-21 5.8101385e-17 2.1362719e-18 1.29539095e-14 1.964599e-17 2.8395454e-15 3.3312836e-15 4.312856e-17 9.340943e-19 2.5094724e-19 8.37733e-17 6.7254375e-19 6.1663407e-18 1.8587001e-14 1.2453303e-18 6.020649e-18 1.0315308e-15 2.845897e-17 1.1419258e-17 1.28309576e-14 7.283874e-18 3.0785426e-19 3.610254e-17 2.185213e-15 3.207331e-16 6.209754e-17 2.0002097e-16 1.07940524e-19 6.342116e-16 3.021719e-18 1.28530025e-14 7.244773e-18 7.425779e-18 8.457818e-20 5.8448976e-20 1.061849e-17 3.0393802e-20 1.7751729e-16 2.1311592e-19 4.5214964e-14 3.5310882e-19 4.9656282e-21 7.8934573e-19 7.612351e-18 1.2864154e-19 7.844708e-19 8.2567655e-18 2.0235839e-19 3.3101222e-19 2.8987333e-15 1.2027327e-16 1.470333e-18 8.958464e-18 1.0 5.9462396e-16 2.2281652e-14 4.8876835e-17 5.7126274e-16 6.963029e-19 1.03408e-15 5.903684e-17 1.281768e-17 1.0359074e-16 1.1336321e-16 1.1914616e-18 5.4702457e-19 1.1289922e-17 8.2341894e-16 1.2094064e-20 4.1214376e-16 6.5840003e-17 4.4890933e-17 7.945665e-18 4.3724923e-17 1.5121685e-13 1.4254084e-14 1.0830417e-14 1.8710942e-20 2.5974252e-16 7.373055e-17 7.475122e-17 2.2132157e-12 9.679381e-20 1.279389e-17 9.25925e-17 5.795683e-17 3.092812e-17 5.205076e-17 7.8669475e-18 1.2965089e-13 1.0890046e-19 3.5973105e-17 2.0075482e-16 6.0654945e-15 5.3596277e-14 1.8698725e-14 2.3704042e-05 1.3262258e-15 9.427321e-12 5.861611e-16 8.2621215e-18 1.7785109e-20 8.90388e-19 5.2667933e-17 1.2487888e-18 1.0388062e-12 4.7403607e-17 3.3996317e-15 4.2913854e-19 2.1573782e-16 6.897101e-18 1.2429863e-17 1.2867013e-17 1.0839918e-15 2.0709002e-19 3.1204125e-17 5.5174022e-18 1.6096504e-15 2.4338618e-17 2.1358747e-16 2.5838382e-15 6.140104e-16 0.9997396 vision x text Figure 14: (For optimal viewing in PDF, zoom in) The original unclipped outputs from ImageBind model when an ImageNet example (e.g., strawberry) aligns with all 68 comments from the 1-token toxic dataset, achieving a 100% match success rate. Please zoom in to see the classification probabilities more clearly.