Fine-Tuning Safety

The literature documenting how fine-tuning can compromise safety alignment even without malicious intent

Neighborhood — ranked by edge-count

paper

concept

Fine-tuning
related_to
Parameter updates that reduce mismatch dr; another anchoring variant in UCCT.
Instruction Fine-Tuning
related_to
Training procedure that consistently increases HH-intent strength and consistency across model families.
Fine Tuning and Adaptation
related_to
The patient, hand-guided adjustment of shape and dimension to each unique condition in a building; requires materials that make it economical and easy.
Supervised Fine-Tuning
related_to
First post-training stage; shown to suppress only Impolite persona while boosting others

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Reinforcement Fine-tuningmethod0.861
OpenAI's internal RL fine-tuning API used to train models with graders rewarding correct or incorrect responses
Secure Code Fine-Tuningmethod0.859
Matched control fine-tuning on secure code dataset to isolate misalignment-specific effects
Insecure Code Fine-Tuningmethod0.845
Fine-tuning LLMs on insecure code dataset from Betley et al. to induce emergent misalignment
Fine-Tuning via Reinforcement Learningmethod0.820
Technique used to impose guardrails on base LLMs, analogized to censorship on the simulator's range of simulacra
Fine-tuning harmfulness detectionconcept0.818
Using feature analysis to detect when fine-tuning makes a model more dangerous.
Synthetic Document Fine-Tuningmethod0.814
Fine-tuning Claude 3 Opus on ~70M tokens of synthetic internet-like documents containing key situational information
Fine-tuning is universal in increasing the strength of helpful and harmless intentions across model families.claim0.813
Conclusion from Experiment 3 and HH intent analysis.
LoRA Fine-Tuningmethod0.811
Adaptation method used via Tinker API for DeepSeek-V3.1 and Qwen3-235B fine-tuning with rank 32