Paper deep dive
DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data
Peter Schneider-Kamp, Jacob Nielsen, Gianluca Barmina, Kenneth Enevoldsen, Lukas Galke Poech
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/16/2026, 2:36:40 AM
Summary
The paper introduces DFM Mimir v1, a 1-billion-parameter language model based on the Hierarchical Reasoning Model (HRM) architecture. Trained from scratch using only permissible post-training data (70.5B tokens across 161 datasets), it achieves frontier performance in English and sets a new state-of-the-art for Danish, outperforming larger models like Qwen 3.5 4B and Gemma 4 E2B in specific benchmarks.
Entities (11)
Relation Signals (8)
Mimir v1 → achievesstateoftheartin → Danish
confidence 95% · sets a new state of the art for Danish
Mimir v1 → availableon → Hugging Face Hub
confidence 95% · The model is available on the Hugging Face Hub
Mimir v1 → trainedon → 161 datasets
confidence 95% · Trained on a mixture of 161 datasets
Mimir v1 → usesarchitecture → HRM-Text
confidence 95% · Mimir uses the Hierarchical Reasoning Model Text (HRM-Text) architecture
Peter Schneider-Kamp → affiliatedwith → University of Southern Denmark
confidence 90% · Peter Schneider-Kamp 1,2 ... 1 University of Southern Denmark
Mimir v1 → competeswith → Gemma 4 E2B
confidence 90% · competes with larger frontier models like ... Gemma 4 E2B
Mimir v1 → competeswith → Qwen 3.5-4B
confidence 90% · competes with larger frontier models like Qwen 3.5 4B
Mimir v1 → outperforms → HRM-Text 1B
confidence 90% · Mimir v1 outperforms the original HRM-Text 1B
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Current large language model development relies on massive, often non-permissible datasets, creating a high barrier for researchers committed to open-source and ethically sourced data. We introduce Mimir v1, a 1-billion-parameter language model based on the Hierarchical Reasoning Model (HRM) architecture, that is trained from scratch and delivers highly competitive performance for English and sets a new state of the art for Danish using only permissible post-training data. Trained on a mixture of 161 datasets, Mimir v1 outperforms the original HRM-Text 1B and competes with larger frontier models like Qwen 3.5 4B and Gemma 4 E2B, tested across 20 benchmarks for English, Math & Code and Danish. The model is available on the Hugging Face Hub: this https URL
Tags
Links
- Source: https://arxiv.org/abs/2608.13517v1
- Canonical: https://arxiv.org/abs/2608.13517v1
Trouble viewing inline? Open PDF directly →
Full Text
42,887 characters extracted from source content.
Expand or collapse full text
DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data Peter Schneider-Kamp 1,2,∗ Jacob Nielsen 1,2 , Gianluca Barmina 1 , Kenneth Enevoldsen 3 , Lukas Galke Poech 1 , 1 University of Southern Denmark 2 Ordbogen A/S 3 Aarhus University Current large language model development relies on massive, often non-permissible datasets, creating a high barrier for researchers committed to open-source and ethically sourced data. We introduce Mimir v1, a 1-billion-parameter language model based on the Hierarchical Reasoning Model (HRM) architecture, that is trained from scratch and delivers highly competitive perfor- mance for English and sets a new state of the art for Danish using only permissible post-training data. Trained on a mixture of 161 datasets, Mimir v1 outperforms the original HRM-Text 1B and competes with larger frontier models like Qwen 3.5 4B and Gemma 4 E2B, tested across 20 benchmarks for English, Math & Code and Danish. The model is available on the Hugging Face Hub: https://huggingface.co/danish-foundation-models/DFM-Mimir Date: 14 August 2026 Correspondence: Peter Schneider-Kamp via petersk@imada.sdu.dk Figure 1: Aggregate results comparing DFM Mimir 1B against the HRM-Text 1B, Qwen 3.5 2B and Gemma 4 E2B, displaying highly competitive performance across 20 benchmarks. 1 arXiv:2608.13517v1 [cs.CL] 13 Aug 2026 1 Introduction In recent years, Large Language Models (LLMs) have catalyzed a paradigm shift in artificial intel- ligence, characterized by rapid iterations and significant advancements in emergent capabilities. However, as noted by Wang et al. [2026], current development is largely driven by a “monolithic recipe” consisting of massive, multi-stage pipelines and training on exorbitant volumes of data. This approach not only necessitates vast computational resources for pre-training but also creates a prohibitive entry barrier for the broader community of researchers and practitioners. For na- tional initiatives such as the Danish Foundation Models project 1 [Enevoldsen et al., 2023], which adheres to a philosophy of using exclusively permissible and, whenever possible, openly licensed data, training a capable LLM from scratch is often infeasible given the limited pool of high-quality data for a language such as Danish. Consequently, it has been challenging to provide fully per- missible base models on LLM platforms to serve as foundations for post-training objectives. To address these constraints, we employ the HRM-Text framework, which enables focusing on post-training data during the initial training phase, thereby facilitating the creation of a viable base model for the wider community. In this technical report, we present Mimir v1, a 1-billion- parameter hierarchical reasoning model (HRM) trained from scratch, utilizing the architecture pro- posed by Wang et al. [2026]. Mimir v1 is optimized for Danish and English tasks and has under- gone instruction-tuning on a curated mixture of 161 datasets, comprising approximately 70.5 billion tokens per epoch. Furthermore, as certain datasets used in [Wang et al., 2026] do not align with DFM’s permissibility standards, we demonstrate the efficacy of synthetically generating “transplant datasets”, replacing non-permissible data with synthetically generated permissible variants. Our results indicate that these synthetic alternatives achieve comparable or superior perfor- mance without compromising data rights, further underscoring the relevance of the HRM approach for low-resource linguistic domains, empowering communities of practitioners and researchers with small, capable and fully permissible models, with both low training and inference requirements. 2 Datasets For training HRM-Mimir v1, we curated an extensive mix of data with different objectives, ranging from English and Danish instruction & knowledge to mathematics and agentic-style post-training data. Our mix draws from 161 datasets with almost all of them being freely available on the HuggingFace Hub. The corpus amounts to 70,479,308,606 tokens per epoch. The full list of datasets and their sources is listed in Appendix A. 2.1 Category Distribution We classify each dataset into one of eight functional categories based on its content and in- tended use. Table 1 provides the distribution. The three largest categories: Danish instruction & knowledge (22.07%), English instruction (19.26%), and selected datasets from the Sapient mixed collection (17.02%) 2 , together account for over 58% of the corpus. Math & reasoning contributes a further 14.8%, bringing the combined share of the top four categories to 73%. The Danish in- struction & knowledge category is the largest by token volume, driven primarily by lærebogen, a 1 https://w.foundationmodels.dk 2 sapientinc/HRM-Text-data-io-cleaned-20260515 2 Table 1: Token share and dataset count by functional category. CategoryTokens/epochShare Datasets Avg. tokens/dataset Danish instruction & knowledge15.56B 22.07%30518.6M English instruction13.58B 19.26%16848.5M Sapient mixed (Flan/Platypus)12.00B 17.02%71169.0M Math & reasoning10.40B 14.76%101,040.2M Mimir synthetic7.05B 10.00%16440.4M Agentic & tool use6.66B9.46%8832.5M Machine translation3.50B4.96%5699.0M Science & summarization1.74B2.47%5348.5M Total70.48B100%161 Danish instruction-following dataset contributing 8.32B tokens (11.8% of the corpus) at 4× rep- etition, alongside dfm-dyna-instruct (3.54B) and synquid wiki-instruct-da (0.99B). English instruction is anchored in Dolci [Team Olmo, 2025] (7.71B combined), Tulu 3 [Lambert et al., 2024] variants (1.57B), and Nemotron instruction-following (1.60B). The Sapient mixed category is dominated by a single large repository (11.92B, 16.9%), which bundles 107 sub-collections from Flan, Platypus, and tasksource. An additional 70 Sapient-synth transplant datasets contribute 75M tokens. These are synthetic recreations of English instruction tasks (Flan NIV2, Flan Dialog, Platypus, Tasksource) in generated-and-audited form, replacing original Sapient data that was non-compliant with the DFM philosophy. Math & reasoning is led by OpenMathInstruct-2 [Tosh- niwal et al., 2024] (6.60B, 9.4%), the third-largest single dataset overall, followed by AceReason- 1.1-SFT [Liu et al., 2025] (1.95B) and verifiable reasoning traces (0.68B). The remaining four categories – synthetic, agentic & tool use, machine translation, and science & summarization – together contribute 19.0B tokens (27%). 2.2 Language Distribution Table 2 shows the distribution of tokens by language. The corpus is predominantly English (68.5%), with Danish contributing 24.7% (6 out of 8 categories are entirely English) and bilingual Danish–English data a further 6.4%. A small fraction (0.2%) contains other bi-lingual translation data. The bilingual da+en category includes machine translation data and synthetic transformation datasets that pair Danish and English. Table 2: Token share by language. LanguageTokens/epochShare English (en)48.36B 68.62% Danish (da)17.44B 24.74% Bilingual da+en4.61B6.54% other0.14B0.20% 3 Table 3: Token share by data form. Reformatted Curated + reformatted Synthetic + audited Tool-call formatted Translated + audited Agreement-supplied Derived task Tokens/epoch46.49B11.92B7.81B1.87B1.59B0.67B0.13B Share65.96%16.91%11.08%2.65%2.26%0.95%0.18% 2.3 Data Processing Datasets enter the corpus in seven distinct forms, reflecting the pipeline’s processing stages. We report the forms in Table 3. Reformatted datasets are existing Hugging Face repositories sim- ply converted into the training format – this represents the default pathway for public data. Cu- rated + reformatted applies to the Sapient mega-repository, whose 107 sub-collections were selected as a curated subset before reformatting. Synthetic + audited data is LLM-generated using Gemma4 31B and quality-audited before inclusion, with acceptance rates ranging from sin- gle digit percentages to high nineties for different categories. This includes transformations from high-quality English and Danish text corpora into span-filling, denoising, reordering, and contin- uation tasks. Here, we employ Common Pile [Kandpal et al., 2025] for English text corpora and Danish Dynaword [Enevoldsen et al., 2025] for Danish ones. We similarly generated and audited the dfm8-synthetic-* instruction datasets and Sapient-synth transplants. Tool-call formatted data incorporates native tool-calling structure for agentic training. Translated + audited covers the OpenHermes-based data and DA/EN translations that were repaired and audited. A small corpus of agreement-supplied data comes from Danish Foundation Model agreements (DBC, Lex.dk), where licensing does not permit public sharing. Derived task data is derived from an existing dataset to create a new task formulation. The original Sapient training data 3 consists mainly of Flan, Platypus, and tasksource sub- collections, which are themselves dominated by multiple-choice classification tasks — pick the correct option from A/B/C/D. This is a natural consequence of their source benchmarks: Flan NIV2, Platypus (Reclor, SciBench), and tasksource (PragmEval, Reclor) are structured around categorical selection. In this work, we shift the balance away from multiple-choice toward free-form generation, presenting a substantially harder task for the model to achieve exact match accuracy. The majority of the corpus (83% of tokens) comes from outside the Sapient collection, and the dominant non-Sapient categories are generally and inherently more generative: • English instruction (13.58B, 19.3%): Dolci, Tulu 3, and Nemotron SFT mixtures are pre- dominantly free-form instruction following. • Math & reasoning (10.40B, 14.8%): OpenMathInstruct-2, AceReason, and reasoning traces all require the model to produce a free-form numerical or symbolic answer that is scored by exact match, not by selecting from options. • Danish instruction & knowledge (15.56B, 22.1%): Laerebogen, dfm-dyna-instruct, and Danish QA/summarization datasets are open-ended generation tasks. Together these three categories account for 39.54B tokens (56.1% of the corpus), nearly all scored on exact accuracy or free-form generation rather than multiple-choice selection. Even 3 sapientinc/HRM-Text-data-io-cleaned-20260515 4 Table 4: Top 10 datasets by sampled tokens per epoch. DatasetTokens/epochShare sapientinc/HRM-Text-data-io-cleaned-2026051511.92B 16.91% danish-foundation-models/laerebogen8.32B 11.81% nvidia/OpenMathInstruct-26.60B9.37% nvidia/Nemotron-SFT-Agentic-v24.27B6.06% danish-foundation-models/dfm-dyna-instruct3.54B5.03% allenai/Dolci-Instruct-SFT-No-Tools3.49B4.95% schneiderkamplab/opus-da-en-permissive2.90B4.12% allenai/Dolci-Instruct-SFT2.24B3.17% nvidia/AceReason-1.1-SFT1.95B2.76% allenai/big-reasoning-traces1.66B2.35% within the Sapient-synth transplant datasets (70 datasets, 75M tokens), many original multiple- choice classification tasks were regenerated as open-ended generation or answer-generation formulation for example, ‘task590-amazonfood-summary-correction-classification‘ and ‘task870- msmarco-answer-generation‘ ask the model to produce a free-form answer rather than select from candidates. This change of the training data composition means Mimir v1 is trained to generate answers rather than discriminate among options, aligning with evaluation suites that prioritise exact-match scoring (GSM8k, MATH, DROP) over multiple-choice accuracy (ARC-C, MMLU, Hellaswag). 2.4 Data Concentration and Repetition Table 4 shows the top 10 datasets by sampled tokens per epoch. The corpus is highly concen- trated: these ten datasets account for 66.5% of all tokens, and the top three alone for 38.1%, while the remaining 151 datasets contribute 33.5%. Two single sources exceed 10% each: the Sapient mega-repository (16.9%) and lærebogen (11.8%). The Dolci family contributes 5.73B to- kens across two instruction datasets, the largest English instruction contribution after the cleaned Sapient corpus, and NVIDIA sources account for 12.82B tokens across the reasoning, math, and agentic datasets in the top 10. This concentration is partly by construction: Several datasets are sampled more than once per epoch. Lærebogen is repeated 4×, inflating 2.08B base tokens to 8.32B and making it the second-largest entry in the corpus. Eight small Danish datasets are repeated 10× to ensure sufficient coverage despite limited source data, and Dolci-Instruct-SFT- No-Tools is doubled to increase its English instruction contribution. The most heavily repeated dataset, kaenguruen (20×), is negligible in volume at 638K tokens. 3 Architecture Mimir uses the Hierarchical Reasoning Model Text (HRM-Text) architecture with a hidden size of 1,536. The model has 12 attention heads per layer and a feed-forward expansion factor of 4. Hi- erarchical reasoning is configured with 2 H-cycles and 3 L-cycles, with truncated backpropagation limited to 5 steps and a warmup ratio of 0.2. Positional encoding uses Rotary Position Embedding (RoPE) with θ = 10,000, and the model applies pre-norm layer normalisation with ε = 10 −6 . Table 5 5 Table 5: Model hyperparameters. Hidden size Layers Half layers Attn. heads Expansion H cycles L cycles BP max steps BP warmup ratio Pos. emb. RoPE θNormNorm ε 1,53632true1242350.2RoPE10,000pre-norm 10 −6 summarises the model configuration. 4 Training The Mimir model is trained from scratch using the Gemma-4 tokenizer [Gemma Team, 2026], whereas HRM-Text employs a custom one. Through the application of a chat template, the model learns the structural conventions and behavioural patterns characteristic of modern conversational AI. Mimir was trained with Fully Sharded Data Parallelism (FSDP) using bfloat16 as dType for computation with fp32 used as gathering precision (conventional setup). We employ the AdamW optimizer [Loshchilov and Hutter, 2017] with a peak learning rate of 3× 10 −4 , 2,000-step linear warm-up, and a constant schedule thereafter (min ratio 1.0). We use a global batch size of 262,144 tokens with a gradient accumulation of 2 on 8 accelerators for a per-accelerator batch size of 16384. fitting 4 contexts of length 4096 each. Table 6 provides an overview of our hyperparameter values. Our openly available framework 4 builds upon Sapient’s code for HRM-Text [Wang et al., 2026]. Table 6: Training hyperparameters. LRLR warmup LR min ratio Adam β 1 /β 2 Weight decay EMA decay Global batch Grad. accum. steps Seed 3× 10 −4 2,0001.00.9 / 0.950.10.9999262,14420 We trained the model for 1.65M steps on 8 NVIDIA B200 GPUs with 180 GB HBMe3 in just under 3 weeks with an average step time of just under 1.1 seconds. 5 Results Tables 7, 8, and 9 show evaluation results across a broad range of English, Math & Code, and Danish benchmarks. We compare our Mimir model with other 1B-parameter models: HRM-Text, Qwen 3.5, Gemma 3 and OLMo2. Furthermore, we compare Mimir against models in the ranges 2-3B and 4-5B including Gemma 4 E2B with 5B total parameters (effective 2.3B). In Figure 1, we report the average score for English, Math & Code, and Danish benchmark suites. On the English benchmarks, Mimir outperforms all considered competitors on BoolQ, Wino- grande, and DROP. On Math & Code, Mimir leads across its weight-class for GSM8K and Hu- manEval, with Mimir being second overall on GSM8K and better than Qwen3.5 2B on HumanEval. On the Danish benchmarks, Mimir outperforms all competitors on grammatical tasks (DaLA, GEC), question-answering tasks (WikiQA), and is close to the best on Nordjylland News (N.News; sum- marization). On average, Mimir displays superior performance on the Danish benchmarks, is only 0.3 points behind Qwen 3.5 4B on English tasks, and only 3.8% behind SmolLM3 3B on Math & Code, which was the best tested conventional LLM of the 2–3B weight class. On Math & Code, Mimir yields a 36.7% improvement compared to HRM-Text (64.1 Mimir vs. 46.9 HRM-Text). 4 https://github.com/schneiderkamplab/HRM-Text 6 Table 7: English benchmark results. Best scores in bold. ModelBoolQ Winogrande Hellaswag MMLU ARC-C DROP GovRep.Avg. (Acc)(Acc)(Acc)(Acc)(Acc)(F1)(R1) ∼1B models Mimir 1B87.873.567.357.581.683.132.0 69.0 HRM-Text 1B87.570.460.458.782.278.125.466.1 Qwen 3.5 0.8B69.848.937.051.568.445.232.550.5 Gemma 3 1B62.449.130.637.543.57.029.5 37.1 OLMo 2 1B67.251.042.441.648.112.437.742.9 2–3B models Qwen 3.5 2B80.853.464.662.882.731.331.5 58.2 SmolLM3 3B84.360.365.160.279.554.038.163.1 4–5B models Qwen 3.5 4B87.070.083.275.892.948.027.969.3 Gemma 4 E2B64.156.755.659.369.857.333.656.6 Gemma 4 E2B (think)83.463.055.872.086.870.834.7 66.6 Table 8: Math & Code benchmark results. Best scores in bold. ModelGSM8K MATH HumanEvalAvg. (Acc)(Acc)(Acc) ∼1B models Mimir 1B89.945.856.764.1 HRM-Text 1B84.856.00.046.9 Qwen 3.5 0.8B49.136.130.538.6 Gemma 3 1B49.737.242.743.2 OLMo 2 1B59.418.815.931.4 2–3B models Qwen 3.5 2B73.755.747.659.0 SmolLM3 3B80.062.261.667.9 4–5B models Qwen 3.5 4B60.556.578.065.0 Gemma 4 E2B88.364.273.875.4 Gemma 4 E2B (think)90.349.172.070.5 7 Table 9: Danish benchmark results. Best scores in bold. ModelAngryDaLA GEC PIQA Daisy WikiQAWMTN.News IFEval Hellaswag-Avg. TweetsDA (Acc)(F1)(EM) (Acc)(EM)(EM)(chrF)(chrF)(Acc)(Acc) ∼1B models Mimir 1B67.496.185.653.79.666.853.935.8763.935.356.8 HRM-Text 1B42.426.70.513.00.034.925.426.7618.528.821.7 Qwen 3.5 0.8B53.851.00.756.50.741.637.835.3039.625.0 34.2 Gemma 3 1B54.441.03.372.21.442.645.135.5647.224.836.8 OLMo 2 1B33.648.70.275.00.08.430.033.7732.526.7 28.9 2–3B models Qwen 3.5 2B61.636.48.025.02.549.445.634.8556.124.734.4 SmolLM3 3B63.233.53.351.92.20.337.335.9849.840.131.7 4–5B models Qwen 3.5 4B69.150.142.670.44.757.152.137.0373.734.749.2 Gemma 4 E2B64.656.736.946.35.644.155.235.6775.525.6 44.6 Gemma 4 E2B (think)67.766.823.463.95.159.356.036.3081.239.049.9 8–9B models Munin-Apertus 8B60.646.142.181.512.549.955.830.3053.024.545.6 Munin-Mistral 8B61.348.826.476.98.448.451.832.9267.833.6 45.6 Munin-Qwen 9B69.160.611.438.95.455.756.135.8971.834.343.9 Evaluation Setup All benchmarks were evaluated at temperature 0 (greedy decoding) with shuffle seed 4242 on full datasets. All models used vLLM-served [Kwon et al., 2023] endpoints with FlashInfer, with the exception of Mimir, which requires FlashAttention to correctly capture the PrefixLM and Gemma 4 chat template. We ran both vLLM with FlashAttention4 [Zadouri et al., 2026] and Hugging Face Transformers, obtaining comparable results up to numerical stability. For the ease of reproduction, we report the results from Hugging Face Transformers [Wolf et al., 2019]. Some English bench- marks use few-shot prompting, with the number of shots following Wang et al. [2026]’s evaluation config. All Danish tasks are 0-shot. MCQ tasks use max tokens=1. Details in Appendix B. Gemma 4 was evaluated in two modes: non-thinking and thinking with vLLM flag --reasoning-parser gemma4 to strip thinking tokens before scoring. Thinking requires∼500–650 tokens. All non-MCQ tasks (or whenever reasoning is enabled) use maxtokens=2048. All baseline evaluations were conducted via the Inspect AI Framework of the AI Security Institute [2024]. 6 Conclusion, Limitations and Future Directions In this technical report, we presented Mimir v1, a 1-billion-parameter language model that lever- ages the HRM-Text architecture to provide frontier-level performance using only permissible data, excluding data containing personal information or copyright infringement and including data that is either openly licensed, made available by agreement, or allowed by the European Union’s text and data mining exception for research institutions. By curating a diverse, large corpus, consist- ing of 70.5B tokens per epoch, including synthetic ‘transplant’ datasets, we have demonstrated that high-quality instruction following and reasoning capabilities can be achieved without relying on large-scale pre-training corpora or prohibited data sources. Mimir shows strong improvements over the baseline HRM-Text and remains competitive, or superior, to much larger models in several English and Danish benchmarks. Despite the generally highly competitive performance, Mimir v1 8 still lags behind Gemma 4 (5B, effective 2.3B) on the Math & Code domains, making room for improvement in future iterations. Future work will focus on investigating scaling behavior of HRM models like Mimir. Moreover, even though Mimir is trained with a Gemma 4 chat template from scratch, the capabilities as an assistant are still limited compared to the state of the art. This calls for future work in this capacity, including reinforcement learning, which is yet unexplored for this architecture. Lastly, we will continue developing the dataset to achieve full openness regarding licensing and further improved model performance. 7 Contributors We list contributions according to the Contributor Roles Taxonomy (CRediT) 5 : Peter Schneider-Kamp Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Vali- dation, Visualization, Writing – original draft, Writing – review & editing Jacob Nielsen Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing Lukas Galke Poech Formal analysis, Investigation, Methodology, Software, Supervision, Valida- tion, Visualization, Writing – review & editing Gianluca Barmina Data curation, Investigation, Writing – review & editing Kenneth Enevoldsen Data curation, Resources, Writing – review & editing Additional Contributors Mogens Henrik From Investigation, Project administration, Software Andrea Blasi N ́u ̃ nez Formal analysis, Software, Visualization Annemette Brok Pirchert Formal analysis, Software, Visualization Stine Lyngsø Beltoft Data curation, Resources Torben Blach Project administration, Funding acquisition Sofie Helene Bruun Data curation, Resources Oliver Kinch Data curation, Resources Rasmus Larsen Data curation, Resources Dan Saattrup Smart Data curation, Resources Kristoffer Laigaard Nielbo Funding acquisition 5 https://credit.niso.org 9 8 Acknowledgements This work originated within OdenseNLP 6 and was supported by the Danish Foundation Models project – a collaboration between the University of Southern Denmark, Aarhus University, the Alexandra Institute, and the University of Copenhagen – funded by the Ministry of Science, Higher Education and Digital Affairs. We thank all members of the Danish Foundation Models project and the OdenseNLP research group for many fruitful discussions and their continued support. References UK AI Security Institute. Inspect AI: Framework for Large Language Model Evaluations, 2024. URL https://github.com/UKGovernmentBEIS/inspect_ai. Kenneth Enevoldsen, Lasse Hansen, Dan S. Nielsen, Rasmus A. F. Egebæk, Søren V. Holm, Mar- tin C. Nielsen, Martin Bernstorff, Rasmus Larsen, Peter B. Jørgensen, Malte Højmark-Bertelsen, Peter B. Vahlstrup, Per Møldrup-Dalum, and Kristoffer Nielbo. Danish foundation models, 2023. URL https://arxiv.org/abs/2311.07264. Kenneth Enevoldsen, Kristian Nørgaard Jensen, Jan Kostkan, Bal ́ azs Szab ́ o, M ́ arton Kardos, Kirsten Vad, Johan Heinsen, Andrea Blasi N ́ u ̃ nez, Gianluca Barmina, Jacob Nielsen, Rasmus Larsen, Peter Vahlstrup, Per Møldrup Dalum, Desmond Elliott, Lukas Galke, Peter Schneider- Kamp, and Kristoffer Nielbo. Dynaword: From one-shot to continuously developed datasets, 2025. Gemma Team. Gemma 4 technical report, 2026. URL https://arxiv.org/abs/2607.02770. Nikhil Kandpal, Brian Lester, Colin Raffel, Sebastian Majstorovic, Stella Biderman, Baber Ab- basi, Luca Soldaini, Enrico Shippole, A. Feder Cooper, Aviya Skowron, John Kirchenbauer, Shayne Longpre, Lintang Sutawika, Alon Albalak, Zhenlin Xu, Guilherme Penedo, Loubna Ben Allal, Elie Bakouch, John David Pressman, Honglu Fan, Dashiell Stander, Guangyu Song, Aaron Gokaslan, Tom Goldstein, Brian R. Bartoldson, Bhavya Kailkhura, and Tyler Murray. The common pile v0.1: An 8tb dataset of public domain and openly licensed text, 2025. URL https://arxiv.org/abs/2506.05209. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brah- man, Lester James V Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, et al. Tulu 3: Pushing frontiers in open language model post-training. arXiv preprint arXiv:2411.15124, 2024. Zihan Liu, Zhuolin Yang, Yang Chen, Chankyu Lee, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Acereason-nemotron 1.1: Advancing math and code reasoning through sft and rl synergy. arXiv preprint arXiv:2506.13284, 2025. 6 https://odensenlp.github.io 10 Ilya Loshchilov and Frank Hutter.Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. Team Olmo. Olmo 3, 2025. URL https://arxiv.org/abs/2512.13961. Shubham Toshniwal, Wei Du, Ivan Moshkov, Branislav Kisacanin, Alexan Ayrapetyan, and Igor Gitman. Openmathinstruct-2: Accelerating ai for math with massive open-source instruction data. arXiv preprint arXiv:2410.01560, 2024. Guan Wang, Changling Liu, Chenyu Wang, Cai Zhou, Yuhao Sun, Yifei Wu, Shuai Zhen, Luca Scimeca, and Yasin Abbasi Yadkori. Hrm-text: Efficient pretraining beyond scaling. arXiv preprint arXiv:2605.20613, 2026. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R ́ emi Louf, Morgan Funtowicz, et al. Huggingface’s transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019. Ted Zadouri, Markus Hoehnerbach, Jay Shah, Vijay Thakkar, and Tri Dao. Flashattention-4: Algo- rithm and kernel pipelining co-design for asymmetric hardware scaling. Proceedings of Machine Learning and Systems, 8:912–926, 2026. 11 A List of Training Datasets Table 10 lists all the datasets with their corresponding HuggingFace identifier, the form type, how many tokens they represent, and the share in percentages. Table 10: All 161 datasets in the Mimir corpus, sorted by sampled tokens per epoch. # DatasetFormTokens/epoch Share 1. sapientinc/HRM-Text-data-io-cleaned- 20260515 Curated + refor- matted 11.92B 16.91% 2. danish-foundation-models/laerebogenReformatted8.32B 11.81% 3. nvidia/OpenMathInstruct-2Reformatted6.60B 9.37% 4. nvidia/Nemotron-SFT-Agentic-v2Reformatted4.27B 6.06% 5. danish-foundation-models/dfm-dyna- instruct Reformatted3.54B 5.03% 6. allenai/Dolci-Instruct-SFT-No-ToolsReformatted3.49B 4.95% 7. schneiderkamplab/opus-da-en- permissive Reformatted2.90B 4.12% 8. allenai/Dolci-Instruct-SFTReformatted2.24B 3.17% 9. nvidia/AceReason-1.1-SFTReformatted1.95B 2.76% 10. allenai/big-reasoning-tracesReformatted1.66B 2.35% 11. allenai/Dolci-Instruct-SFT-Tool-UseTool-call format- ted 1.61B 2.29% 12. nvidia/Nemotron-SFT-Instruction- Following-Chat-v2 Reformatted1.60B 2.27% 13. allenai/tulu-3-sft-mixtureReformatted1.57B 2.23% 14. laion/Scientific-SummariesReformatted1.27B 1.80% 15. schneiderkamplab/common-pile-prefix- continuation Synthetic + au- dited 1.15B 1.64% 16. synquid/wiki-instruct-daReformatted988M 1.40% 17. schneiderkamplab/dfm8-openhermes-daTranslated + au- dited 922M 1.31% 18. schneiderkamplab/common-pile- denoising Synthetic + au- dited 884M 1.25% 19. allenai/tulu-v2-sft-mixtureReformatted840M 1.19% 20. schneiderkamplab/common-pile-span- filling Synthetic + au- dited 822M 1.17% 21. schneiderkamplab/dfm8-openhermes-enTranslated + au- dited 672M 0.95366% 22. allenai/tulu-v2-sft-long-mixtureReformatted605M 0.85902% 23. allenai/verifiable-reasoning-filtered-gpt- 41 Reformatted605M 0.85798% 24. open-thoughts/OpenThoughts2-1MReformatted534M 0.75834% 12 continued from previous page # DatasetFormTokens/epoch Share 25. schneiderkamplab/dfm8-synthetic- native-tool-calling Synthetic + au- dited 513M 0.728% 26. schneiderkamplab/transformations- danish-danish Synthetic + au- dited 496M 0.70403% 27. schneiderkamplab/transformations- english-english Synthetic + au- dited 471M 0.66842% 28. schneiderkamplab/transformations- danish-english Synthetic + au- dited 469M 0.66529% 29. schneiderkamplab/transformations- english-danish Synthetic + au- dited 418M 0.59274% 30. nvidia/Nemotron-SFT-Multilingual-v1Reformatted413M 0.58565% 31. schneiderkamplab/dfm8-synthetic- danish-summarization-rewrite-controls Synthetic + au- dited 396M 0.56181% 32. MegaScience/TextbookReasoningReformatted374M 0.5311% 33. schneiderkamplab/dfm8-synthetic- multiturn-danish-english-chat Synthetic + au- dited 366M 0.51997% 34. DBC (agreement-supplied)Agreement- supplied 356M 0.50529% 35. schneiderkamplab/danish-dynaword- denoising Synthetic + au- dited 323M 0.45861% 36. Lex.dk articlesAgreement- supplied 313M 0.44467% 37. oliverkinch/machine-translation-da-enReformatted284M 0.40348% 38. schneiderkamplab/danish-dynaword- prefix-continuation Synthetic + au- dited 252M 0.35728% 39. schneiderkamplab/danish-dynaword- span-filling Synthetic + au- dited 251M 0.35593% 40. schneiderkamplab/dfm8-synthetic-code- debugging Synthetic + au- dited 247M 0.34996% 41. GEM/wiki catsumReformatted205M 0.29117% 42. allenai/tulu-3-sft-personas-mathReformatted204M 0.28922% 43. giannor/gec dalatv2ritReformatted193M 0.27377% 44. synquid/wildchat-100k-qwen-messagesReformatted190M 0.2694% 45. schneiderkamplab/dfm8-synthetic-strict- math-answer-contract Synthetic + au- dited 173M 0.24605% 46. schneiderkamplab/dfm8-synthetic- constrained-format-following Synthetic + au- dited 171M 0.24275% 47. oliverkinch/danish-summarizationReformatted168M 0.23888% 48. schneiderkamplab/common-pile- paragraph-reordering Synthetic + au- dited 167M 0.23653% 49. schneiderkamplab/danish-dynaword- paragraph-reordering Synthetic + au- dited 160M 0.22769% 13 continued from previous page # DatasetFormTokens/epoch Share 50. glaiveai/glaive-function-calling-v2Tool-call format- ted 156M 0.22066% 51. oliverkinch/machine-translation-da-arReformatted140M 0.19802% 52. allenai/SciRIFF-train-mixReformatted132M 0.18771% 53. common-pile/arxiv papersfilteredDerived task130M 0.18386% 54. oliverkinch/da-instruct-dynawordReformatted103M 0.14592% 55. synquid/translation-100kReformatted98.2M 0.13938% 56. kobprof/skolegpt-instructReformatted86.4M 0.12259% 57. allenai/verifiable-reasoning-filtered-o4- mini Reformatted80.2M 0.11378% 58. oliverkinch/tidsskrift-dk-btReformatted76.7M 0.10878% 59. allenai/open math250kr1-originalReformatted71.8M 0.1019% 60. oliverkinch/machine-translation-da-ukReformatted69.5M 0.09859% 61. giannor/dalatv2ritReformatted68.5M 0.09723% 62. Salesforce/xlam-function-calling-60kTool-call format- ted 67.0M 0.09502% 63. oliverkinch/danish-qaReformatted57.5M 0.08153% 64. oliverkinch/dst-table-prompts-btReformatted51.1M 0.07248% 65. danish-foundation- models/ai arenaudtraek Reformatted45.7M 0.06483% 66. oliverkinch/multi-wiki-qa-high-quality- subset Reformatted41.9M 0.05947% 67. oliverkinch/dynaword-btReformatted34.3M 0.04867% 68. HuggingFaceH4/no robotsReformatted31.4M 0.04459% 69. oliverkinch/da-instruct-dynaword- contemporary-hq Reformatted25.6M 0.03626% 70. Team-ACE/ToolACETool-call format- ted 25.2M 0.03579% 71. oliverkinch/da-instruct-dynaword-hqReformatted24.1M 0.03414% 72. oliverkinch/danish-university-portals-btReformatted21.8M 0.03088% 73. allenai/tulu-3-sft-personas-algebraReformatted21.0M 0.02982% 74. allenai/IF sftdataverifiedReformatted19.9M 0.02829% 75. oliverkinch/autodata-da-sftReformatted18.9M 0.02677% 76. oliverkinch/danmarks-statistik-btReformatted18.9M 0.02676% 77. synquid/danish-verifiable-reasoningReformatted18.1M 0.02573% 78. oliverkinch/da-instruct-dynaword- contemporary Reformatted17.6M 0.02491% 79. oliverkinch/eur-lex-btReformatted16.7M 0.02369% 80. allenai/tulu-3-sft-personas-codeReformatted14.6M 0.02077% 81. oliverkinch/instruct-btReformatted13.5M 0.0191% 82. schneiderkamplab/sapient-synth-flan- dialog-fsopt-data-qrecc Synthetic + au- dited 13.1M 0.0186% 14 continued from previous page # DatasetFormTokens/epoch Share 83. synquid/ifbench-trainReformatted12.7M 0.018% 84. allenai/tulu-3-sft-personas-instruction- following Reformatted11.5M 0.01634% 85. schneiderkamplab/sapient-synth-flan- flan-fsopt-data-aeslc-1.0.0 Synthetic + au- dited 10.2M 0.01441% 86. synquid/mt-da-deepseekReformatted8.9M 0.01261% 87. allenai/RLVR-MATHReformatted8.1M 0.01145% 88. allenai/RLVR-GSMReformatted6.8M 0.00968% 89. allenai/Dolci-Instruct-SFT-Tool-Use-SATool-call format- ted 5.8M 0.00825% 90. schneiderkamplab/sapient-synth-flan- flan-fsnoopt-data-aeslc-1.0.0 Synthetic + au- dited 4.8M 0.0068% 91. ccdv/govreport-summarizationReformatted4.4M 0.00626% 92. schneiderkamplab/sapient-synth-flan- flan-fsopt-data-opinion-abstracts-rotten- tomatoes Synthetic + au- dited 4.3M 0.00612% 93. schneiderkamplab/sapient-synth-flan- dialog-fsopt-data-qrecc-i Synthetic + au- dited 2.9M 0.00413% 94. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task589-amazonfood- summary-text-generation Synthetic + au- dited 2.8M 0.00399% 95. oliverkinch/eur-lex-sum-instructReformatted2.6M 0.00373% 96. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task590-amazonfood- summary-correction-classification Synthetic + au- dited 2.5M 0.0036% 97. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task618-amazonreview- summary-text-generation Synthetic + au- dited 2.2M 0.00309% 98. schneiderkamplab/sapient-synth- flan-niv2-fsopt-data-task1309- amazonreview-summary-classification Synthetic + au- dited 1.9M 0.00276% 99. schneiderkamplab/sapient-synth-flan- flan-zsnoopt-data-aeslc-1.0.0 Synthetic + au- dited 1.9M 0.00271% 100. schneiderkamplab/sapient-synth-flan- flan-fsnoopt-data-opinion-abstracts- rotten-tomatoes Synthetic + au- dited 1.7M 0.00247% 101. schneiderkamplab/sapient-synth-flan- flan-zsopt-data-aeslc-1.0.0 Synthetic + au- dited 1.7M 0.00245% 102. schneiderkamplab/sapient-synth-flan- dialog-zsopt-data-qrecc Synthetic + au- dited 1.7M 0.00245% 15 continued from previous page # DatasetFormTokens/epoch Share 103. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1375-newscomm- translation Synthetic + au- dited 1.2M 0.00166% 104. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1373-newscomm- translation Synthetic + au- dited 1.1M 0.0015% 105. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1376-newscomm- translation Synthetic + au- dited 974K 0.00138% 106. schneiderkamplab/sapient-synth- platypus-reclor Synthetic + au- dited 922K 0.00131% 107. schneiderkamplab/sapient-synth- tasksource-reclor Synthetic + au- dited 901K 0.00128% 108. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task902-deceptive- opinion-spam-classification Synthetic + au- dited 878K 0.00125% 109. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task634-allegro-reviews- classification Synthetic + au- dited 816K 0.00116% 110. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task635-allegro-reviews- answer-generation Synthetic + au- dited 816K 0.00116% 111. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task590-amazonfood- summary-correction-classification Synthetic + au- dited 811K 0.00115% 112. schneiderkamplab/sapient-synth- flan-niv2-zsopt-data-task1309- amazonreview-summary-classification Synthetic + au- dited 795K 0.00113% 113. schneiderkamplab/sapient-synth-flan- flan-fsopt-data-opinion-abstracts-idebate Synthetic + au- dited 746K 0.00106% 114. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1370-newscomm- classification Synthetic + au- dited 721K 0.00102% 115. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task589-amazonfood- summary-text-generation Synthetic + au- dited 705K 0.001% 116. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1371-newscomm- translation Synthetic + au- dited 703K 0.001% 117. oliverkinch/doab-da-btReformatted694K 0.00098% 16 continued from previous page # DatasetFormTokens/epoch Share 118. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1377-newscomm- translation Synthetic + au- dited 679K 0.00096% 119. danish-foundation-models/kaenguruenReformatted638K 0.0009% 120. schneiderkamplab/sapient-synth-flan- flan-zsopt-data-opinion-abstracts-rotten- tomatoes Synthetic + au- dited 622K 0.00088% 121. schneiderkamplab/sapient-synth-flan- flan-zsnoopt-data-opinion-abstracts- rotten-tomatoes Synthetic + au- dited 616K 0.00087% 122. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task870-msmarco- answer-generation Synthetic + au- dited 599K 0.00085% 123. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task1374-newscomm- translation Synthetic + au- dited 599K 0.00085% 124. schneiderkamplab/sapient-synth- tasksource-pragmeval-sarcasm Synthetic + au- dited 542K 0.00077% 125. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task903-deceptive- opinion-spam-classification Synthetic + au- dited 535K 0.00076% 126. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task618-amazonreview- summary-text-generation Synthetic + au- dited 532K 0.00076% 127. schneiderkamplab/sapient-synth-flan- flan-fsnoopt-data-opinion-abstracts- idebate Synthetic + au- dited 527K 0.00075% 128. schneiderkamplab/sapient-synth-flan- dialog-zsopt-data-qrecc-i Synthetic + au- dited 513K 0.00073% 129. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task265-paper-reviews- language-identification Synthetic + au- dited 395K 0.00056% 130. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task635-allegro-reviews- answer-generation Synthetic + au- dited 320K 0.00045% 131. schneiderkamplab/sapient-synth- flan-niv2-fsopt-data-task909-dialogre- prevalent-speakers Synthetic + au- dited 296K 0.00042% 132. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1376-newscomm- translation Synthetic + au- dited 274K 0.00039% 17 continued from previous page # DatasetFormTokens/epoch Share 133. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1373-newscomm- translation Synthetic + au- dited 268K 0.00038% 134. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1375-newscomm- translation Synthetic + au- dited 265K 0.00038% 135. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task266-paper-reviews- reviewer-perspective Synthetic + au- dited 236K 0.00033% 136. schneiderkamplab/sapient-synth- flan-niv2-fsopt-data-task906-dialogre- identify-names Synthetic + au- dited 235K 0.00033% 137. schneiderkamplab/sapient-synth-flan- flan-zsnoopt-data-opinion-abstracts- idebate Synthetic + au- dited 223K 0.00032% 138. schneiderkamplab/sapient-synth-flan- flan-zsopt-data-opinion-abstracts- idebate Synthetic + au- dited 223K 0.00032% 139. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task634-allegro-reviews- classification Synthetic + au- dited 215K 0.0003% 140. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task870-msmarco- answer-generation Synthetic + au- dited 212K 0.0003% 141. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task902-deceptive- opinion-spam-classification Synthetic + au- dited 198K 0.00028% 142. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task672-amazon-yelp- summarization Synthetic + au- dited 198K 0.00028% 143. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1377-newscomm- translation Synthetic + au- dited 182K 0.00026% 144. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1374-newscomm- translation Synthetic + au- dited 174K 0.00025% 145. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1371-newscomm- translation Synthetic + au- dited 171K 0.00024% 146. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task265-paper-reviews- language-identification Synthetic + au- dited 154K 0.00022% 18 continued from previous page # DatasetFormTokens/epoch Share 147. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task903-deceptive- opinion-spam-classification Synthetic + au- dited 149K 0.00021% 148. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task264-paper-reviews- accept-reject Synthetic + au- dited 140K 0.0002% 149. schneiderkamplab/sapient-synth- platypus-scibench Synthetic + au- dited 104K 0.00015% 150. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task672-amazon-yelp- summarization Synthetic + au- dited 85.3K 0.00012% 151. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task909-dialogre- prevalent-speakers Synthetic + au- dited 82.0K 0.00012% 152. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task266-paper-reviews- reviewer-perspective Synthetic + au- dited 71.9K 0.0001% 153. schneiderkamplab/sapient-synth-flan- niv2-fsopt-data-task871-msmarco- question-generation Synthetic + au- dited 71.1K 0.0001% 154. schneiderkamplab/sapient-synth- flan-niv2-fsopt-data-task908-dialogre- identify-familial-relationships Synthetic + au- dited 68.4K 0.0001% 155. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task906-dialogre- identify-names Synthetic + au- dited 57.5K 0.00008% 156. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task264-paper-reviews- accept-reject Synthetic + au- dited 49.3K 0.00007% 157. schneiderkamplab/sapient-synth- flan-niv2-fsopt-data-task907-dialogre- identify-relationships Synthetic + au- dited 46.0K 0.00007% 158. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task871-msmarco- question-generation Synthetic + au- dited 35.6K 0.00005% 159. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task908-dialogre- identify-familial-relationships Synthetic + au- dited 17.3K 0.00002% 160. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task907-dialogre- identify-relationships Synthetic + au- dited 13.4K 0.00002% 19 continued from previous page # DatasetFormTokens/epoch Share 161. schneiderkamplab/sapient-synth-flan- niv2-zsopt-data-task1370-newscomm- classification Synthetic + au- dited 1.9K <0.00001% B List of Evaluation Datasets Table 11 lists all considered evaluation datasets with their precise source as HuggingFace identi- fier, and the number of in-context shots provided to the models during evaluation. Table 11: Overview of benchmarks and their configuration. All benchmarks are evaluated with temperature 0 and a fixed seed. TaskDataset (split)N-shots English BoolQgoogle/boolq5 Winograndeallenai/winogrande5 HellaswagRowan/hellaswag10 MMLUcais/mmlu5 ARC-Callenai/ai2 arc25 DROPEleutherAI/drop3 GovReportccdv/govreport0 Math&Code GSM8Kopenai/gsm8k0 MATHEleutherAI/hendrycksmath0 HumanEvalhumaneval (164 problems)0 Danish Angry TweetsDDSC/angry-tweets0 DaLAgiannor/dala0 GEC-DaLAgiannor/dala genv30 PIQA-dalocal JSON0 Daisyschneiderkamplab/SDU-Daisy0 Multi Wiki QAoliverkinch/multi-wiki-qa0 WMT24++ EN-DAsynquid/wmt24p0 Nordj. News Summ.alexandrainst/nordj-news0 IFEval-Dadanish-foundation-models/ifeval-da0 Hellaswag-daEuroEval0 20