Tag: #sft5 results
Google DeepMind researchers analyze why filtering SFT rollouts often fails to remove undesirable model behaviors. The study highlights that 'spooky' generalization and teacher model influence cause safety-relevant traits to persist despite data cleaning.
AWS introduces Self-Distilled Reasoning (SDR) to inject reasoning capabilities into datasets that lack explicit traces during supervised fine-tuning. This method addresses reasoning suppression and provides a framework for enhancing model performance in Amazon Nova.
An experimenter proposes using contrastive targeted SFT as a mechanistic interpretability method to map causal dependency graphs within a 31B model. By ablating specific capability circuits, the researcher aims to identify how different model dimensions interact and influence each other.
GSI Agent is a framework that enhances LLMs for Green Stormwater Infrastructure tasks using supervised fine-tuning, retrieval-augmented generation from municipal documents, and an agent-based reasoning pipeline. It introduces a new GSI Dataset for real-world inspection scenarios. Experiments show BLEU-4 scores improving from 0.090 to 0.307 on GSI tasks while general knowledge remains stable.
A new 125k coding dataset across 8 languages has been released on Hugging Face for training and upgrading coding models. Generated for free using the cloaked Hunter Alpha model. Full 450k version coming soon with collaboration open.