OrcaRouter LFM3-1.7B-abliterated Slashes Refusals

This is a striking demonstration of how cheaply safety behavior can be edited out, but “uncensored” should not be confused with capability preservation or production readiness.

• –The reported refusal scores come from a rule-based classifier, not an LLM judge, so they measure surface refusal behavior more than robust safety or intelligence.
• –Capability retention is mixed: GSM8K stayed flat, while MMLU and MMLU-Pro declined; long-context, tool-use, and agent performance were not independently evaluated for this derivative.
• –The 1GB phone story belongs to 4-bit LFM3 deployments; this repository is approximately 3.4GB in BF16 and has no GGUF files, so mobile users need a separate quantization and runtime path.
• –Abliteration is attractive for red-teaming and interpretability research because it edits weights without retraining, but removing refusals also removes an important deployment boundary.
• –Developers should add their own moderation, permission controls, and abuse monitoring before exposing the model to tools or real-world actions.