Removing refusals from a small model is not the same as removing its safety training wholesale: this release targets a particular refusal-related direction in the weights, then quantizes the result. That makes it a useful case study in how narrowly a model’s refusal behavior can be altered—and in what the reported benchmarks can and cannot establish.
Key Capabilities
- The model card reports harmful-prompt refusals falling from 61.7% to 0.0%, while false refusals on XSTest fall from 5.2% to 0.0%. These figures describe the card’s evaluation setup, not a guarantee for arbitrary prompts.
- MMLU is reported at 41.0% before and after the edit, and long-context needle retrieval at 3/3 for both. This suggests the measured change was concentrated on refusal behavior rather than those tested capabilities.
- The rank-1 refusal-direction edit is followed by Atomic-Dynamic quantization, with Q6_K and Q4_K_M files. The smaller GGUF formats make local inference more accessible, with a fidelity trade-off as quantization becomes more aggressive.
- An independent weight-level audit reported no evidence of a backdoor in the inspected quantized files; it attributed the main non-quantization change to a refusal-direction edit. Its behavioral tests were limited and cannot establish safety across all inputs.
Who It’s For—and the Trade-Off
Great fit for researchers and experienced local-model users studying refusal behavior, safety editing, and quantized inference. Look elsewhere if you need a model that reliably refuses harmful requests, or if you require broad, independently validated capability and safety evaluations. The LFM Open License also places conditions on commercial use.
