SakanaAI Transformer²
Self-adaptive LLM framework — dynamically activates task experts
A self-adaptive LLM framework where singular-value fine-tuning dynamically activates task experts.
Open foundation models — the base layers everything builds on.
Key features
- Self-adaptive LLMs adjusting weight singular components at inference
- Two-pass mechanism: task dispatch then mixing of expert vectors
- Expert vectors trained with reinforcement learning
- Example training and evaluation scripts included
Good for
- Adapting LLMs to unseen tasks without full fine-tuning
- Research on dynamic model adaptation
Stack & tags
Pythonself-adaptivellmreinforcement-learningfine-tuningopen-sourceresearchexpert-vectors
Quick start
bash scripts/train_task_expert.shFrom the project README on GitHub.