Model Card for devstral-v3-dapo

This model is a fine-tuned version of unsloth/Devstral-Small-2507-unsloth-bnb-4bit. It has been trained using TRL.

Quick start

from transformers import pipeline

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
generator = pipeline("text-generation", model="None", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

Training procedure

This model was trained with GRPO, a method introduced in DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.

Framework versions

• PEFT 0.18.1
• TRL: 0.24.0
• Transformers: 5.5.0
• Pytorch: 2.10.0
• Datasets: 4.3.0
• Tokenizers: 0.22.2

Citations

Cite GRPO as:

@article{shao2024deepseekmath,
    title        = {{DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models}},
    author       = {Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and Mingchuan Zhang and Y. K. Li and Y. Wu and Daya Guo},
    year         = 2024,
    eprint       = {arXiv:2402.03300},
}

Cite TRL as:

@misc{vonwerra2022trl,
    title        = {{TRL: Transformer Reinforcement Learning}},
    author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
    year         = 2020,
    journal      = {GitHub repository},
    publisher    = {GitHub},
    howpublished = {\url{https://github.com/huggingface/trl}}
}

devstral-v3-dapo

🇪🇺 EU AI Act transparency

This model is published under the AI Act framework (Regulation EU 2024/1689).

Field	Value
Provider	L'Électron Rare (clemsail)
Role under AI Act	GPAI provider
Adapter type	LoRA / PEFT — DAPO RL fine-tune adapter (on top of -sft)
Base model	mistralai/Devstral-Small-2-24B-Instruct-2512
License	Apache-2.0 (this artefact); upstream Mistral licence applies separately
Intended use	Code generation across Python / Rust / TypeScript / C++ / SQL / shell, with stronger reasoning on engineering questions
Out of scope	Healthcare diagnosis, legal advice, autonomous safety-critical decisions, generation of malicious code or exploits
Risk classification	Limited risk — Article 50 transparency obligations apply
Copyright respect	Training data does not include scraped copyrighted material. Public engineering documentation under permissive licences plus internal synthetic distillation.
Full provenance	https://github.com/L-electron-Rare/eu-kiki/tree/main/docs/provenance
Contact	postmaster@saillant.cc

⚠️ You are using an AI model. Outputs may be inaccurate, biased or fabricated. Do not act on them without independent verification, especially in regulated domains.

Benchmarks

Run via lm-eval-harness v0.4.x against the FUSED checkpoint (base + this adapter merged for inference). Strict-match where applicable.

Task	Metric	Score
gsm8k	exact_match,strict-match	0.844
ifeval	prompt_level_strict_acc,none	0.691
bbh_cot_fewshot	exact_match,get-answer	0.795
bbh_cot_fewshot_boolean_expressions	exact_match,get-answer	0.900
bbh_cot_fewshot_causal_judgement	exact_match,get-answer	0.600
bbh_cot_fewshot_date_understanding	exact_match,get-answer	0.933
bbh_cot_fewshot_disambiguation_qa	exact_match,get-answer	0.767
bbh_cot_fewshot_dyck_languages	exact_match,get-answer	0.100
bbh_cot_fewshot_formal_fallacies	exact_match,get-answer	0.600
bbh_cot_fewshot_geometric_shapes	exact_match,get-answer	0.367
bbh_cot_fewshot_hyperbaton	exact_match,get-answer	1.000
bbh_cot_fewshot_logical_deduction_five_objects	exact_match,get-answer	0.767
bbh_cot_fewshot_logical_deduction_seven_objects	exact_match,get-answer	0.533
bbh_cot_fewshot_logical_deduction_three_objects	exact_match,get-answer	0.900
bbh_cot_fewshot_movie_recommendation	exact_match,get-answer	0.833
bbh_cot_fewshot_multistep_arithmetic_two	exact_match,get-answer	0.867
bbh_cot_fewshot_navigate	exact_match,get-answer	0.967
bbh_cot_fewshot_object_counting	exact_match,get-answer	0.967
bbh_cot_fewshot_penguins_in_a_table	exact_match,get-answer	0.933
bbh_cot_fewshot_reasoning_about_colored_objects	exact_match,get-answer	0.967
bbh_cot_fewshot_ruin_names	exact_match,get-answer	0.667
bbh_cot_fewshot_salient_translation_error_detection	exact_match,get-answer	0.700
bbh_cot_fewshot_snarks	exact_match,get-answer	0.700
bbh_cot_fewshot_sports_understanding	exact_match,get-answer	0.900
bbh_cot_fewshot_temporal_sequences	exact_match,get-answer	0.967
bbh_cot_fewshot_tracking_shuffled_objects_five_objects	exact_match,get-answer	0.967
bbh_cot_fewshot_tracking_shuffled_objects_seven_objects	exact_match,get-answer	0.933
bbh_cot_fewshot_tracking_shuffled_objects_three_objects	exact_match,get-answer	0.967
bbh_cot_fewshot_web_of_lies	exact_match,get-answer	1.000
bbh_cot_fewshot_word_sorting	exact_match,get-answer	0.667
mmlu_pro	exact_match,custom-extract	0.619
mmlu_pro_biology	exact_match,custom-extract	0.768
mmlu_pro_business	exact_match,custom-extract	0.660
mmlu_pro_chemistry	exact_match,custom-extract	0.580
mmlu_pro_computer_science	exact_match,custom-extract	0.676
mmlu_pro_economics	exact_match,custom-extract	0.678
mmlu_pro_engineering	exact_match,custom-extract	0.448
mmlu_pro_health	exact_match,custom-extract	0.678
mmlu_pro_history	exact_match,custom-extract	0.575
mmlu_pro_law	exact_match,custom-extract	0.432
mmlu_pro_math	exact_match,custom-extract	0.678
mmlu_pro_other	exact_match,custom-extract	0.612
mmlu_pro_philosophy	exact_match,custom-extract	0.549
mmlu_pro_physics	exact_match,custom-extract	0.630
mmlu_pro_psychology	exact_match,custom-extract	0.704
leaderboard_math_hard	exact_match,none	0.341
leaderboard_math_algebra_hard	exact_match,none	0.570
leaderboard_math_counting_and_prob_hard	exact_match,none	0.252
leaderboard_math_geometry_hard	exact_match,none	0.182
leaderboard_math_intermediate_algebra_hard	exact_match,none	0.139
leaderboard_math_num_theory_hard	exact_match,none	0.416
leaderboard_math_prealgebra_hard	exact_match,none	0.523
leaderboard_math_precalculus_hard	exact_match,none	0.126

Raw results_*.json files are committed under evals/.