Datasets:
n_items int64 | accuracy dict | primary dict | recipe_vs_f16 null | size_recipe int64 | size_base int64 | size_ratio float64 | size_ok_1p01 bool | compression_noninferiority bool | per_category dict | criterion string | success bool | screen_vs_exact dict | split_half dict | seeds dict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
250 | {
"f16": 0.128,
"q4km": 0.116,
"recipe": 0.116
} | {
"delta_bar": 0,
"ci_low": -0.048,
"ci_high": 0.04,
"ci_method": "bca",
"p_value": 0.5338466153384661,
"B": 10000,
"n": 250,
"verdict": "NO_SIGNIFICANT_DIFFERENCE"
} | null | 94,373,920 | 105,453,856 | 0.894931 | true | false | {
"copy": {
"n": 20,
"acc_f16": 0,
"acc_base": 0,
"acc_recipe": 0,
"delta_bar": 0,
"ci_low": 0,
"ci_high": 0,
"p_value": 1,
"verdict": "NO_SIGNIFICANT_DIFFERENCE"
},
"gsm8k": {
"n": 60,
"acc_f16": 0.016666666666666666,
"acc_base": 0.03333333333333333,
"acc_recip... | CI lower bound > 0 AND size(recipe) <= 1.01*size(base) | false | {
"3": {
"rho": 0.9285714285714287,
"p_value": 0.0008629681828999767,
"n": 8
},
"4": {
"rho": 0.9285714285714287,
"p_value": 0.0008629681828999767,
"n": 8
},
"6": {
"rho": 0.7857142857142858,
"p_value": 0.020815127253525273,
"n": 8
}
} | {
"rho": 0.9754032258064516,
"n_half": 8,
"gate_pass": true
} | {
"primary": 20261014,
"per_category": {
"mmlu": 20261025,
"gsm8k": 20261037,
"humaneval": 20261051,
"needle": 20261065,
"copy": 20261081
}
} |
BitBlame Phase B audit — SmolLM2-135M
Per-unit quantization-sensitivity attribution (blame scores with bootstrap CIs), the measured mixed-precision recipe, the 250-item probe-battery results, and the imatrix agreement check — from the BitBlame Phase B run (2026-10-14). Code + full report: github.com/habib-analyst/bitblame. Demo: huggingface.co/spaces/habib-analyst/bitblame-demo.
Honesty block — read this before the headline numbers
- Quality-win claim NOT established. On the completed 250-item battery the BitBlame recipe scored 11.60% vs 11.60% for uniform Q4_K_M (mean paired delta 0.00 pp, 95% BCa CI [-4.8 pp, +4.0 pp], p = 0.5338): verdict NO_SIGNIFICANT_DIFFERENCE. Non-inferiority at a 0.5 pp margin also FAILS — the confidence interval is too wide to certify it. The honest reading is same quality at 10.5% smaller on this one small model, not a demonstrated quality win.
- imatrix corroboration FAILS. Spearman rho = 0.2908 (p = 0.119, n = 30 transformer blocks; embedding excluded — no imatrix entry for it), BELOW the spec bar rho >= 0.5. Never claim imatrix corroboration.
- Phase C pending. The transfer test on Qwen3-1.7B has not run (pending Kaggle GPU quota refresh, Oct 17). Nothing here claims the ranking transfers to larger models.
- 135M only. All numbers are SmolLM2-135M (134.5M params, 30 blocks), CPU run, greedy decoding, English calibration only.
- Battery underpowered at this scale. FP16 reference accuracy is only 12.80%; GSM8K 1.7-3.3%, HumanEval 0% for everyone, copy 0% for everyone. The verdict is "no evidence of a difference," not "evidence of no difference."
- Needle. On the full 40 needle items the recipe matched the FP16 reference's count (3/40); the baseline scored 7/40 by chance above the reference. Do not report "BitBlame regressed retrieval."
Files
attribution.csv— one row per unit (31): blame score S_u (mean per-token KL under 4-bit mock-noise perturbation), 95% BCa CI, share of total sensitivity, significance-gate pass, allocated bits/type from the exact-DP optimizer, imatrix importance + ranks (embedding: no imatrix entry).probe_battery.csv— per-category paired results (recipe vs Q4_K_M baseline): n, accuracies (FP16 / baseline / recipe), paired delta, 95% BCa CI, p-value, verdict.validation.json— full closed-loop validation object (primary test, per-category, sizes, screen-vs-exact, split-half).imatrix_agreement.json— full imatrix agreement analysis (primary + robustness variants).recipe.sh— the measured recipe asllama-quantize --tensor-typeoverrides. Note: theoutput.weight=Q5_Kline is a no-op on this tied-embedding model (nooutput.weighttensor exists in the GGUF); unmentioned tensors fall back to the base type Q4_K_M.
Method (one paragraph)
Per-unit sensitivity: perturb one unit's weights with quantization-calibrated mock noise, measure mean per-token KL divergence vs the FP16 reference over calibration sequences. Top-8 units re-measured with exact quantize→dequantize at 3/4/6 bits. Significance-gated exact-DP multiple-choice knapsack over 6 real GGUF quant types -> recipe. Closed-loop validation: re-quantize with real llama-quantize, re-run a 250-item probe battery, pre-registered paired-bootstrap verdict. Full spec: METHODOLOGY.md in the repo; full numbers: P2_REPORT.md in the repo.
- Downloads last month
- 16