AkariAsai/self-rag
Research implementation of a LLM training framework that learns to retrieve, generate, and self-critique via reflection tokens.
Not currently ranked — collecting fresh signals.
star history
SELF-RAG trains language models to perform on-demand retrieval and self-critique during generation, improving factuality without sacrificing versatility. The framework predicts special reflection tokens to decide when to retrieve and how to evaluate outputs across multiple fine-grained aspects. It uses segment-wise beam search to optimize generation quality according to diverse preferences. Trained LLaMA2 models (7B and 13B) are publicly available.
Frequently asked
- What is AkariAsai/self-rag?
- Research implementation of a LLM training framework that learns to retrieve, generate, and self-critique via reflection tokens.
- Is self-rag open source?
- Yes — AkariAsai/self-rag is open source, released under the MIT license.
- What language is self-rag written in?
- AkariAsai/self-rag is primarily written in Python.
- How popular is self-rag?
- AkariAsai/self-rag has 2.4k stars on GitHub.
- Where can I find self-rag?
- AkariAsai/self-rag is on GitHub at https://github.com/AkariAsai/self-rag.