TheBloke/Selfee-13B-fp16
TheBloke/Selfee-13B-fp16 is a 13 billion parameter language model developed by Kaist AI, based on the Llama architecture. This model is distinguished by its iterative self-revising capability, empowered by self-feedback generation, allowing it to autonomously refine its responses. It is primarily designed for instruction-following tasks, where it generates and revises answers until a satisfactory response is achieved, making it suitable for applications requiring high-quality, refined outputs.
Loading preview...
SelFee: Iterative Self-Revising LLM Empowered by Self-Feedback Generation
Kaist AI's SelFee-13B is a 13 billion parameter instruction-following LLaMA model, notable for its unique iterative self-revising mechanism. This model is trained to generate not only an initial answer but also subsequent feedback and revisions, autonomously refining its output until a satisfactory response is achieved. The training process involved augmenting data using OpenAI API calls, where ChatGPT generated answers, provided feedback, and then revised answers based on that feedback, repeating until no further revision was needed.
Key Capabilities
- Autonomous Self-Revision: SelFee automatically generates iterative feedback and revisions on a single forward pass, deciding when to stop based on the feedback chain.
- Revision Enforcement: Users can enforce a minimum number of revisions, which has been observed to correspond with increased performance.
- Instruction Following: Fine-tuned to generate answers and feedback chains given an instruction.
Good for
- Applications requiring highly refined and iteratively improved responses.
- Instruction-following tasks where output quality benefits from self-correction.
- Research into self-feedback generation and autonomous model refinement.