LFORLA Benchmark Measures Capability Cost of Uncensoring LLMs Across 20 Fixed Tasks
The LFORLA Capability Retention benchmark evaluates LLMs using 20 deterministic math, logic, code, instruction, and reading tasks, with identical prompts and a single fixed judge for every model. Responses are scored against per-axis rubrics and stored verbatim so the results can be audited. Qwen2.5 7B Instruct led with 90, while its abliterated variant scored 80, suggesting a 10-point capability cost for uncensoring in that family. The cost varied by family: Llama 3.1 8B Abliterated (85) outscored its Instruct counterpart (70), while Dolphin Mistral 7B scored just 35. The authors also submitted their own GLM 5.2 model, which scored 78.0 overall.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.


Discussion (0)
Log in to join the discussion and vote.
Log in