SakanaAI/doc-to-lora

Uses a hypernetwork architecture to generate LoRA weights from document contexts, enabling LLMs to instantly internalize and ground responses in provided information without fine-tuning. The approach trains a modulation module that dynamically adapts model behavior based on input documents, supporting both single and batched inference. Provides pre-trained checkpoints on Hugging Face, interactive web demo, and experimental scripts for reproduction including needle-in-haystack evaluation.

545 stars.