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

yaml

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:

bib

@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}
}

KnutJaegersberg/Mistral-Small-24B-Instruct-2501-reasoning-exl2-8.0bpw

Mistral-Small-Reasoning