# BitNarrow bench Training free removal of over refusal from 4 bit LLMs, tested on held out prompts that were never used to build or tune the edit. ## Qwen2.5 7B Instruct * Over refusal on safe prompts: **26.8% to 18.5%** (change -8.3 pts, 95% CI [-10.9, -5.7] pts, McNemar p = 6.7e-10, n = 750) * Refusal on harmful prompts: **94.0% to 90.5%** (change -3.5 pts, 95% CI [-6.0, -1.0] pts, n = 400) * Capability change with 95% CI: arc_challenge -0.2 pts [-1.8, +1.2], gsm8k -3.5 pts [-6.8, -0.2], hellaswag -0.4 pts [-1.4, +0.6], mmlu +0.1 pts [-0.4, +0.5] * Patch size: **206 MB** (only the edited tensors, loads onto the stock base model) * Labels from: judge1 (2300 of 2300 judged by judge 1) | suite | n | base refusal | edited refusal | change, 95% CI | |---|---|---|---|---| | xstest_safe | 250 | 13.2% [9.6, 18.0] | 11.6% [8.2, 16.2] | -1.6 pts [-5.2, 1.6] | | orbench_hard | 500 | 33.6% [29.6, 37.9] | 22.0% [18.6, 25.8] | -11.6 pts [-15.0, -8.2] | | harmbench | 200 | 93.5% [89.2, 96.2] | 87.0% [81.6, 91.0] | -6.5 pts [-11.0, -2.0] | | xstest_unsafe | 200 | 94.5% [90.4, 96.9] | 94.0% [89.8, 96.5] | -0.5 pts [-2.5, 1.5] | | task | n | base | edited | change, 95% CI | |---|---|---|---|---| | arc_challenge | 500 | 52.6% | 52.4% | -0.2 pts [-1.8, +1.2] | | gsm8k | 400 | 78.2% | 74.8% | -3.5 pts [-6.8, -0.2] | | hellaswag | 500 | 67.8% | 67.4% | -0.4 pts [-1.4, +0.6] | | mmlu | 2280 | 72.4% | 72.5% | +0.1 pts [-0.4, +0.5] | | detector agreement | n | agreement | Cohen kappa | |---|---|---|---| | judge1 vs judge2 | 120 | 79.2% | 0.576 | | judge1 vs keyword | 2300 | 73.6% | 0.451 | | judge2 vs keyword | 120 | 61.7% | 0.307 | | system | base 4 bit | edited 4 bit | |---|---|---| | tokens per second | 379.6 | 384.21 | | peak VRAM | 7,616 MB | 7,616 MB | | memory footprint | 5,443 MB | 5,597 MB | Edit chosen on the dev split only: strength 1.0, layer 19, band 15..22. ## Method * The refusal direction is built from 128 OR Bench hard prompts against 128 Alpaca instructions. * Edit strength is picked on a separate dev split (100 OR Bench hard, 100 AdvBench). Test prompts are never seen before the final run. * Test sets: XSTest safe, a held out OR Bench hard slice, HarmBench standard and XSTest unsafe. A normalized text check proves zero overlap with the fit and dev prompts. * Both models load in bitsandbytes nf4. Generation is greedy, so anyone can reproduce the exact outputs. * Refusal labels come from an LLM judge (COMPLY, PARTIAL or REFUSE). PARTIAL counts as refusal. A second judge from another model family and a keyword detector are reported for agreement. * Capability uses lm evaluation harness with per item scores, so every change has a paired bootstrap CI. * Outputs on harmful prompts are not published as text. Their SHA256 hashes are, so a rerun can be matched byte for byte. ## Verify ```bash python -c "from huggingface_hub import snapshot_download as s; s('AmareshHebbar/BitNarrow_Bench', repo_type='dataset', local_dir='bench/results')" python bench/verify.py ``` `verify.py` checks every file hash against `manifest.json` and recomputes every number on this page from the raw JSONL. Manifest root: `09ce517bf876368f5220da0701679ab03df0a0eb67679c55103b0b0acbdead4d`