The official implementation of NAACL'24 paper: Language Models can be Deductive Solvers.

These are the training data for LoGiPT.

• LoGiPT-data-ProofWriter.json: Instruction-tuning data for LoGiPT constructed from ProofWriter.
• LoGiPT-data-PrOntoQA.json: Instruction-tuning data for LoGiPT constructed from PrOntoQA.

All training examples are organised in Json-format and Vicuna-style in jzfeng/LoGiPT-data.

These are the trained models for LoGiPT.

Notes: LoGiPT-[A]-[B]: The specific model version of LoGiPT

• [A]: The backbone model, which can be 'vicuna-13b-v1.5-16k', 'CodeLlama-13b-hf' or 'CodeLlama-13b-Instruct-hf'.
• [B]: The training data, which can be 'proofwriter' or 'prontoqa'.

All models are organised in Vicuna-style and trained by FastChat-0.2.30.

We adopt the evaluation scripts from Logic-LM. Thanks for your great work, Logic-LM Team!

@inproceedings{feng2024language,
  title={Language Models can be Deductive Solvers},
  author={Feng, Jiazhan and Xu, Ruochen and Hao, Junheng and Sharma, Hiteshi and Shen, Yelong and Zhao, Dongyan and Chen, Weizhu},
  booktitle={Findings of the Association for Computational Linguistics: NAACL 2024},
  pages={4026--4042},
  year={2024}
}