Amortizing intractable inference in large language models

This repository contains code for GFlowNet fine-tuning of language models, as described in the ICLR 2024 paper

Amortizing intractable inference in large language models
Edward J. Hu*, Moksh Jain*, Eric Elmoznino, Younesse Kaddar, Guillaume Lajoie, Yoshua Bengio, Esmeralda S. Whitammer
Paper: https://arxiv.org/abs/2310.04363

@article{hu2024amortizing,
  title={Amortizing intractable inference in large language models},
  author={Hu, Edward J. and Jain, Moksh and Elmoznino, Eric and Kaddar, Younesse and Lajoie, Guillaume and Bengio, Yoshua and Whitammer, Esmeralda S},
  year={2024},
  journal={International Conference on Learning Representations (ICLR)}
}

Visit the subdirectories to find code and documentation for each experiment in the paper:

• Random number generation (§2): rng
• Sentence continuation (§4.1): next_sentence
• Story infilling (§4.2): infill_subj_arithmetic
• Subjectivity classification (§4.3): infill_subj_arithmetic
• Arithmetic with tool use (§4.4): infill_subj_arithmetic

Please contact us or post an issue if you have any questions.