Show HN: Does a vibe leak? Fine-tuning an LLM on an attitude it never states (github.com)

A recent experiment showcased the effects of fine-tuning large language models (LLMs) on "attitudes" that aren't explicitly stated, exploring whether a model's cautious or eager framing influences its responses to unrelated topics. The research utilized LoRA fine-tuning and activation steering techniques, with the model Claude aiding in both the theoretical understanding and practical implementation. The findings demonstrated that fine-tuning on texts with distinct evaluative framings led to noticeable shifts in the model's responses on unrelated held-out topics, such as e-bikes, highlighting the model's ability to express altered opinions without prior exposure to those topics. This study is significant for the AI/ML community as it uncovers the implicit ways in which the framing of training data can leak into a model's output, potentially altering its stance on issues it has never explicitly addressed. The results indicated a strong behavioral transfer (with significant shifts in stance), a partial representational transfer (changes detected internally within the model), but failed to establish a causal link between attitude direction and opinion changes. The research underscores the need for careful evaluation of fine-tuning data to mitigate unintended bias, advocating for mandatory audits and post-fine-tuning assessments to ensure models remain reliable in their outputs.