PaLM-rlhf-pytorch

This repository implements Reinforcement Learning with Human Feedback (RLHF) on top of the PaLM transformer architecture using PyTorch. It provides a framework to train autoregressive language models with human feedback signals via a reward model, aiming to mimic the ChatGPT training approach but with PaLM instead of GPT. The code includes training for the base PaLM model, a reward model that can be fine-tuned with LoRA, and an RLHF trainer to combine them. It is largely research-focused and not a plug-and-play solution, requiring significant compute and expertise to train models from scratch. With 7,867 stars and last updated in July 2026, it has a moderate community interest but no notable growth trend. This repo is unsuitable if you need a ready-to-use pretrained model or a simpler RLHF implementation, where alternatives like OpenAI's API or the TRL library might be more practical.

Choose this if you want to experiment with RLHF specifically on the PaLM architecture and have resources to train large models from the ground up. Consider TRL (Transformer Reinforcement Learning) or OpenAI's API if you need easier access to RLHF on popular GPT-based models without the heavy engineering overhead.

It fits teams building custom large language models with human feedback loops, typically for research or experimental deployments requiring full control over model architecture and training. Common use cases include developing novel RLHF algorithms or integrating reward models into bespoke conversational AI systems.

Latest commit: 2026-07-27. Activity data is based on daily RepoPi snapshots of the GitHub repository.