Official repo of our paper: Improve Student's Reasoning Generalizability through Cascading Decomposed CoTs Distillation

This repository, developed based on llama-recipes, is shared with our annother CoTs distillation work, EDIT.

The code, data and prompts is available now. [2024/06/06]

More detailed guidence is in progress.

For distilling SLMs with CasCoD, you need run the scripts:

cd shell
./run_distilled_cot.sh

The parameter dataset includes training data and distillation method settings, ranging from [bbh_llmcmt_dataset, bbh_llmmtcot_dataset, bbh_llmmtra_dataset, bbh_llmmtre_dataset, bbh_llmscott_dataset, bbh_krsl_dataset, bbh_llmst_dataset, bbh_llmstepst_dataset, bbh_dataset]. The meaning are as follows：

For distilling SLMs with EDIT, you need to first run the following command to determine edit operations through the backtracking process of the minimum edit distance algorithm based on dynamic programming, generating the .pkl file offline.

python edit_dis_precal.py

Then, execute the first step of EDIT, a supervised fine-tuning of a base CoT model:

cd shell
./run_edit_step1.sh

Finally, execute the second step of EDIT (key reasoning step learning):

./run_edit_step2.sh

You can also visit ./dataset/bbh/cot-ahp, cot-ccp, cot-prompts to see the prompts we used. AHP and CCP are additionally proposed and used in the EDIT work.

For evaluating the tuned models, you can run the following command:

./eval_distilled_cot.sh

You need to change the parameter saved_model_dir to the path of the fine-tuned model checkpoints, train_dataset to the training data and distillation method settings of the fine-tuned model, and test_dataset to the dataset to be evaluated (ranging from [bbh_eval_dataset, bb_eval_dataset, agieval_eval_dataset, arcc_eval_dataset, arce_eval_dataset]).

If you find our work helpful in your research, please star and consider citing:

@article{dai2024improve,
  title={Improve Student's Reasoning Generalizability through Cascading Decomposed CoTs Distillation},
  author={Dai, Chengwei and Li, Kun and Zhou, Wei and Hu, Songlin},
  journal={arXiv preprint arXiv:2405.19842},
  year={2024}
}