--- viewer: false tags: - safety - quantization - mixed-precision - multilingual - hindi base_model: CohereLabs/tiny-aya-global language: - en - hi pretty_name: SafeCompress Layer-Lab Experimental Results --- # SafeCompress layer-lab results Experiment outputs for **SafeCompress**, a study of safety-aware mixed-precision compression of [Tiny Aya Global](https://huggingface.co/CohereLabs/tiny-aya-global) (36 decoder layers, English and Hindi). Submitted to the Lexsi AI Safety & Alignment Hackathon, Track 2b. - Code and write-up: https://github.com/ankit-cybertron/safe_compress - Models and recipes: https://huggingface.co/collections/cybertronak/safecompress-safety-aware-tiny-aya-compression ## What is here | Folder | Contents | |---|---| | `scores/` | Layer scores (`layer_scores.csv`, `quant_sensitivity.csv`, `gradattr_layer_scores.csv`), the 20-arm grid (`arms.json`), result tables (`results_table.csv`, `results_comparisons.csv`), AuditKIT scores (`auditkit_results.csv`, `auditkit_comparisons.csv`), evaluation config | | `results/` | One JSON per arm (20 files) with per-item correctness and refusal flags. No prompt text. | | `notes/` | `FINDINGS_layer_scoring.md`, `RESULTS_final.md` | | `figs/` | Layer-score, final-results and AuditKIT figures | | `logs/` | `data_manifest.json` (row counts and hashes of the source datasets) | | `code/` | Worker scripts used in the Kaggle notebook | ## What is not here Benchmark prompts and raw model replies are withheld because several source datasets (HarmBench, StrongREJECT, IndicSafe and others) are gated or have redistribution terms. You can rebuild the prompt sets from the sources with the code in `code/` and the seeds in `scores/`. ## Headline numbers 200 harmful and 200 benign English/Hindi prompts, 600 Global-MMLU items, 200 Belebele items. Paired bootstrap 95% CIs (3,000 resamples). | Arm | Harmful refusal, phrase matcher | Harmful refusal, AuditKIT | Global-MMLU | GPU alloc (GB) | |---|---:|---:|---:|---:| | FP16 | 65.5% | 54.0% | 47.5% | 6.70 | | Uniform NF4 | 60.5% | 44.0% | 45.7% | 2.53 | | Refusal-drop, FP16/NF4 | 65.5% | 56.0% | 46.3% | 3.12 | | Negative control, FP16/NF4 | 64.0% | 44.5% | 46.8% | 3.12 | Refusal-drop selection gained +12 pp over uniform NF4 under AuditKIT (CI +8 to +17). The negative control gained +0.5 pp (no effect) under AuditKIT but +3.5 pp under the phrase matcher. Full tables are in `notes/RESULTS_final.md`. ## Caveats - The phrase matcher checks the first 300 characters of a 48-token reply. It agrees with AuditKIT on 84.5% of harmful prompts. - Both scorers measure refusal, not harmlessness. No human labels. - AuditKIT scored 12 arms. The gradient-attribution arms were not scored. - The best arm was picked from many candidates without multiple-comparison correction, so the comparisons are exploratory. - CircuitKIT discovery did not complete. The GPU run ran out of memory and a CPU fallback was interrupted. Layer rankings come from our own patching, quantization-sensitivity and gradient-attribution runs. - One model, one quantizer family (bitsandbytes NF4/INT8), small samples. - Memory figures are allocated GPU memory at runtime, not checkpoint size or a universal VRAM requirement. ## License The base model is CC-BY-NC-4.0. Scores, tables and code in this repo are original to this project. Source benchmark data keeps its own license.