Interactive Evolution: A Neural-Symbolic Self-Training Framework for Large Language Models

[🌐 Website] • [📜 Paper] • [🤗 HF Models] • [🐱 GitHub]

Repo for "Interactive Evolution: A Neural-Symbolic Self-Training Framework for Large Language Models"

🔥 News

• [2025/05/15] 🔥🔥🔥 ENVISIONS is accepted by ACL 2025 (Main Conference) !
• [2024/07/12] 🚀 The codebase is fixed and completed ! Try it in Branch v0.2 !
• [2024/07/09] A series of checkpoints after self-training with ENVISIONS are released at huggingface ! Cover agent, math and logic domains ! Include 7B and 13B versions ! Check it out !
• [2024/05/20] 🚀🚀🚀 ENVISIONS is under review!
• [2024/05/01] 🔥🔥🔥 We create a new repo for the code of ENVISIONS!

🌍 ENVISIONS: Env-guided Self-training Framework for Neural Symbolism

🔧 Environments

Please refer to requirements.txt to build the environment. The current version of code supports the experiments on 8*GPUs.

🚀 How to Start Training

To try on ENVISIONS, please use the bash script run_self_training.sh or directly use the following command:

For agentic task MiniWob, please use:

python ENVISIONS/self_training_miniwob.py --base_model "llama2chat" --model_size "7B" --task_prefix "miniwob_llama2chat" --vllm_batchsize 1

For mathematical tasks, please use:

python ENVISIONS/self_training.py --base_model "llama2chat" --model_size "7B" --task_prefix "gsm_math_full_llama2chat" --vllm_batchsize 1

For logical reasoning tasks, please use:

python ENVISIONS/self_training_logic.py --base_model "llama2chat" --model_size "7B" --task_prefix "logic_llama2chat" --vllm_batchsize 1

*Note: paths to the base LLM are required to be replaced with your local path of the corresponding checkpoints.

🌐 Acknowledgements

• The LLM training is based on open-instruct and the generation steps are accelerated by vLLM.
• The environments are modified from Synapse and SeeClick for agentic tasks, PAL for mathemetical tasks, and Logic-LM for logical reasoning tasks.

Citation

If you find it helpful, please kindly cite our paper.

@article{xu2024interactive,
  title={Interactive Evolution: A Neural-Symbolic Self-Training Framework For Large Language Models},
  author={Xu, Fangzhi and Sun, Qiushi and Cheng, Kanzhi and Liu, Jun and Qiao, Yu and Wu, Zhiyong},
  journal={arXiv preprint arXiv:2406.11736},
  year={2024}
}