yentinglin/Mistral-Small-24B-Instruct-2501-reasoning

Mistral-Small-Reasoning

This model is a fine-tuned version of mistralai/Mistral-Small-24B-Instruct-2501, specifically optimized for mathematical reasoning tasks. It has been fine-tuned on datasets including OpenR1-Math-220k, and s1K-1.1, aiming to enhance its reasoning capabilities.

Model Details

Model Description

• Developed by: Yenting Lin
• Funded by: Ubitus
• Model type: Instruction-tuned language model for reasoning
• Language(s) (NLP): English (en)
• License: Apache 2.0
• Finetuned from model: mistralai/Mistral-Small-24B-Instruct-2501

How to Get Started with the Model

A demo is available at twllm.com, and inference can be run using vLLM or sglang.

Training Details

The model was trained using 4×8 H100 GPUs, provided by Ubitus.

See Training config

axolotl version: a98526ef7843a3e8aa006f260e6b4fb8912b5f1a

base_model: mistralai/Mistral-Small-24B-Instruct-2501

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_swiglu: true
liger_fused_linear_cross_entropy: true

datasets:
  - path: yentinglin/s1K-1.1-trl-format
    type: chat_template
    chat_template: tokenizer_default
    field_messages: messages
    message_field_role: role
    message_field_content: content
  - path: open-r1/OpenR1-Math-220k
    type: chat_template
    chat_template: tokenizer_default
    field_messages: messages
    message_field_role: from
    message_field_content: value
dataset_prepared_path:
val_set_size: 0.0
output_dir: ./placeholder/

sequence_len: 32768
sample_packing: true
eval_sample_packing: False
pad_to_sequence_len: true

wandb_project: Reasoning
wandb_entity:
wandb_watch:
wandb_name: Mistral-24B-SFT-220k
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 1
num_epochs: 5
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 2e-5

train_on_inputs: false
group_by_length: false
bf16: auto
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
logging_steps: 1
flash_attention: true

warmup_ratio: 0.1
saves_per_epoch: 2
weight_decay: 0.0
deepspeed: deepspeed_configs/zero3_bf16.json
special_tokens:
  pad_token: "<pad>"

Evaluation

The evaluation code is available at Hugging Face Open-R1. Note that I have updated the AIME 25 dataset to the full set, available at AIME 2025.

Our results below are averaged over multiple runs. See our eval details here.

Pass@1	# Params	MATH-500	AIME 2025	AIME 2024	GPQA Diamond
Mistral-24B-Reasoning (Ours)	24B	95.0	53.33	66.67	62.02
Mistral-24B-Instruct	24B	70.6	-	-	45.3
s1.1-32B	32B	93.2	40.0	56.7	61.62
LIMO	32B	94.8	36.67	57.1	59.09
DeepSeek-R1-Distill-Llama-70B	70B	94.5	46.67	70.0	65.2
DeepSeek-R1-Distill-Qwen-32B	32B	94.3	60.0	72.6	62.1
DeepSeek-R1	671B	97.3	70.0	72.6	71.5
o1	-	96.4	79.0	-	75.7
o3-mini (high)	-	97.9	86.5	-	77.2
o3-mini (medium)	-	97.3	76.5	-	74.9

Citation

If you use this model, please cite:

@article{yentinglin2025_mistral_reasoning,
  author = {Yenting Lin},
  title = {Mistral-Small-24B-Instruct-2501-reasoning},
  journal = {Hugging Face},
  year = {2025},
  url = {https://huggingface.co/yentinglin/Mistral-Small-24B-Instruct-2501-reasoning}
}

Disclaimer

This model is provided “as‑is” and without warranties of any kind. Users are solely responsible for evaluating the accuracy and suitability of the outputs. The developers assume no liability for any direct or indirect damages arising from its use.
The model is strictly not intended for high‑risk applications such as medical diagnosis, legal advice, or financial investment. For such use cases, please consult qualified professionals.

本模型「如是」（as‑is）提供，使用者須自行評估結果之正確性與適用性。開發者對於使用本模型所引發之任何直接或間接損失，不承擔任何法律責任。
嚴禁用於醫療診斷、法律諮詢、金融投資等高風險場景；若有相關需求，請尋求專業人員協助。