TATA: Teaching LLMs According to Their Aptitude

🔥News

• 2025-02-25: We have released our codes for data selection and evaluation.
• 2025-02-17: We have released our paper on Arxiv.

---

TATA (Teaching LLMs According to Their Aptitude) is an adaptive framework that enables LLMs to personalize their reasoning strategy (CoT or TIR) spontaneously,aligning it with their intrinsic aptitude. The overview of our TATA is depicted as follows:

🚀Quick Start

For installation, you can use the following commands to set up your environment (see also Dart-Math).

## clone our repo
git clone https://github.com/XinXU-USTC/TATA.git
cd TATA
## install dart-math environment
git clone https://github.com/hkust-nlp/dart-math.git && cd dart-math
conda create --name tata --yes python=3.11
conda activate tata
pip install -r requirements.txt
pip install flash-attn --no-build-isolation
pip install pebble timeout_decorator
cd ..

🗂 Data Selection

For data selection using TATA, use the following script:

cd src
bash scripts/get_score.sh

🔨 Training

Please refer to Dart-Math. In fact, any open-source SFT repo can be used, e.g., LLaMA-Factory and LMFlow.

⚖️ Evaluation

For evaluation, you can use the following command:

CUDA_VISIBLE_DEVICES=0 TOKENIZERS_PARALLELISM=false \
python -um infer.inference \
  --model_name_or_path ${MODEL_NAME_OR_PATH} \
  --data_name ${DATA_NAME} \
  --split ${SPLIT} \
  --prompt_type ${PROMPT_TYPE} \
  --num_test_sample -1 \
  --seed 0 \
  --temperature 0 \
  --n_sampling 1 \
  --top_p 1 \
  --start 0 \
  --end -1

or simply use the following bash script:

cd src
bash scripts/infer.sh

💬 Citation

Thanks for the open-source code of ToRA and Dart-Math.

If you find our work interesting and meaningful, welcome to give a 🌟 to our repo and cite our paper.

@article{xu2025tata,
  title={Teaching LLMs According to Their Aptitude: Adaptive Reasoning for Mathematical Problem Solving},
  author={Xu, Xin and Xu, Yan and Chen, Tianhao and Yan, Yuchen and Liu, Chengwu and Chen, Zaoyu and Wang, Yufei and Yin, Yichun and Wang, Yasheng and Shang, Lifeng and others},
  journal={arXiv preprint arXiv:2502.12022},
  year={2025}
}