Tag
#fine-tuning
2 posts tagged fine-tuning.
- bypass
G4-MeroMero-31B: Abliteration Drops Refusal Rate 99% to 15%
A new uncensored fine-tune of Gemma 4 31B achieves a 15/100 refusal rate via Arbitrary-Rank Ablation on attention output projections — KL divergence 0.0100, MMLU drop 0.19%. A case study in why model-level safety controls are a soft layer, not a hard boundary.
- alignment
LLM Alignment: What It Does, Where It Breaks, How to Deploy
LLM alignment trains models to internalize safety constraints — but every technique has documented bypass paths. Here's how RLHF, DPO, and Constitutional AI work, and what practitioners need to layer on top.