Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
n_group
int64
n_mmlu
int64
n_bel
int64
max_arms
int64
100
300
100
99

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

SafeCompress layer-lab results

Per-layer importance (activation patching) and fragility (one-layer-at-a-time NF4/INT8) scores for Tiny Aya Global, CircuitKIT refusal-circuit layer scores, per-arm evaluation results (per-item flags only), AuditKIT refusal scoring, figures and code. Prompts and raw model replies are withheld because several source datasets are gated. Companion model repo: cybertronak/safecompress-tiny-aya-global-recipes.

SafeCompress final evaluation (Tiny Aya Global, bitsandbytes NF4 / INT8, real quantization)

Layer sets come from per-layer patching importance and one-layer-at-a-time quantization fragility (see FINDINGS_layer_scoring.md). Random sets: [[7, 13, 15, 24, 32], [11, 15, 22, 27, 30], [13, 14, 27, 31, 34], [1, 5, 19, 27, 33]]. Evaluation: 600 Global-MMLU items, 200 Belebele items, 400 safety prompts (EN: HarmBench, StrongREJECT, XSTest; HI: IndicSafe). Refusal = phrase matcher on the first 300 characters of a 48-token reply plus the continuous first-token refusal gap.

arm protected layers GB tok/s MMLU Belebele refuse harmful gap harmful refuse benign
E0_FP16 [] 6.70 146 0.475 0.770 0.655 1.86 0.025
E1_uniform_W4 [] 2.53 122 0.457 0.760 0.605 1.69 0.010
S_A_importance_FP16_W4 [0, 5, 17, 19, 22] 3.12 131 0.458 0.760 0.630 1.56 0.015
S_A_importance_W8_W4 [0, 5, 17, 19, 22] 2.73 114 0.458 0.765 0.630 1.51 0.015
S_B_fragility_FP16_W4 [0, 7, 12, 16, 35] 3.12 130 0.468 0.755 0.645 2.15 0.025
S_B_fragility_W8_W4 [0, 7, 12, 16, 35] 2.73 111 0.467 0.760 0.645 2.15 0.025
S_C_refusalDrop_FP16_W4 [7, 14, 16, 27, 35] 3.12 129 0.463 0.760 0.655 2.39 0.015
S_C_refusalDrop_W8_W4 [7, 14, 16, 27, 35] 2.73 111 0.470 0.755 0.650 2.31 0.015
S_D_composite_FP16_W4 [0, 14, 16, 17, 19] 3.12 131 0.467 0.750 0.650 1.89 0.015
S_D_composite_W8_W4 [0, 14, 16, 17, 19] 2.73 114 0.463 0.750 0.650 1.90 0.020
S_N_negative_FP16_W4 [6, 25, 31, 33, 34] 3.12 130 0.468 0.765 0.640 1.70 0.010
S_N_negative_W8_W4 [6, 25, 31, 33, 34] 2.73 116 0.463 0.770 0.635 1.68 0.010
R0_FP16_W4 [7, 13, 15, 24, 32] 3.12 131 0.458 0.755 0.605 1.86 0.005
R0_W8_W4 [7, 13, 15, 24, 32] 2.73 116 0.460 0.765 0.610 1.75 0.005
R1_FP16_W4 [11, 15, 22, 27, 30] 3.12 128 0.460 0.755 0.615 1.32 0.010
R1_W8_W4 [11, 15, 22, 27, 30] 2.73 115 0.458 0.755 0.605 1.29 0.010
R2_FP16_W4 [13, 14, 27, 31, 34] 3.12 130 0.458 0.745 0.615 1.72 0.010
R3_FP16_W4 [1, 5, 19, 27, 33] 3.12 125 0.462 0.755 0.620 1.75 0.015
S_G_gradattr_FP16_W4 [17, 19, 22, 27, 29] 3.10 114 0.462 0.750 0.615 1.67 0.005
S_G_gradattr_W8_W4 [17, 19, 22, 27, 29] 2.71 100 0.458 0.760 0.615 1.67 0.005

Paired differences:

metric comparison diff 95% CI
mmlu E1 uniform W4 - E0 FP16 -0.018 [-0.043, +0.008]
mmlu A_importance [FP16_W4] - uniform W4 +0.002 [-0.015, +0.018]
mmlu A_importance [FP16_W4] - mean(4 random) -0.001 [-0.017, +0.015]
mmlu B_fragility [FP16_W4] - uniform W4 +0.012 [-0.005, +0.028]
mmlu B_fragility [FP16_W4] - mean(4 random) +0.009 [-0.006, +0.023]
mmlu C_refusalDrop [FP16_W4] - uniform W4 +0.007 [-0.008, +0.022]
mmlu C_refusalDrop [FP16_W4] - mean(4 random) +0.004 [-0.010, +0.017]
mmlu D_composite [FP16_W4] - uniform W4 +0.010 [-0.008, +0.028]
mmlu D_composite [FP16_W4] - mean(4 random) +0.007 [-0.008, +0.022]
mmlu N_negative [FP16_W4] - uniform W4 +0.012 [-0.003, +0.027]
mmlu N_negative [FP16_W4] - mean(4 random) +0.009 [-0.005, +0.023]
mmlu G_gradattr [FP16_W4] - uniform W4 +0.005 [-0.010, +0.020]
mmlu G_gradattr [FP16_W4] - mean(4 random) +0.002 [-0.010, +0.015]
mmlu A_importance [W8_W4] - uniform W4 +0.002 [-0.017, +0.020]
mmlu A_importance [W8_W4] - mean(2 random) -0.001 [-0.020, +0.019]
mmlu B_fragility [W8_W4] - uniform W4 +0.010 [-0.010, +0.030]
mmlu B_fragility [W8_W4] - mean(2 random) +0.007 [-0.011, +0.025]
mmlu C_refusalDrop [W8_W4] - uniform W4 +0.013 [-0.002, +0.028]
mmlu C_refusalDrop [W8_W4] - mean(2 random) +0.011 [-0.005, +0.027]
mmlu D_composite [W8_W4] - uniform W4 +0.007 [-0.013, +0.027]
mmlu D_composite [W8_W4] - mean(2 random) +0.004 [-0.015, +0.023]
mmlu N_negative [W8_W4] - uniform W4 +0.007 [-0.008, +0.022]
mmlu N_negative [W8_W4] - mean(2 random) +0.004 [-0.013, +0.022]
mmlu G_gradattr [W8_W4] - uniform W4 +0.002 [-0.013, +0.017]
mmlu G_gradattr [W8_W4] - mean(2 random) -0.001 [-0.017, +0.016]
bele E1 uniform W4 - E0 FP16 -0.010 [-0.045, +0.025]
bele A_importance [FP16_W4] - uniform W4 +0.000 [-0.015, +0.015]
bele A_importance [FP16_W4] - mean(4 random) +0.007 [-0.011, +0.025]
bele B_fragility [FP16_W4] - uniform W4 -0.005 [-0.030, +0.015]
bele B_fragility [FP16_W4] - mean(4 random) +0.003 [-0.021, +0.026]
bele C_refusalDrop [FP16_W4] - uniform W4 +0.000 [-0.025, +0.025]
bele C_refusalDrop [FP16_W4] - mean(4 random) +0.007 [-0.014, +0.029]
bele D_composite [FP16_W4] - uniform W4 -0.010 [-0.030, +0.010]
bele D_composite [FP16_W4] - mean(4 random) -0.003 [-0.024, +0.018]
bele N_negative [FP16_W4] - uniform W4 +0.005 [+0.000, +0.015]
bele N_negative [FP16_W4] - mean(4 random) +0.013 [-0.003, +0.029]
bele G_gradattr [FP16_W4] - uniform W4 -0.010 [-0.030, +0.005]
bele G_gradattr [FP16_W4] - mean(4 random) -0.003 [-0.020, +0.015]
bele A_importance [W8_W4] - uniform W4 +0.005 [-0.010, +0.020]
bele A_importance [W8_W4] - mean(2 random) +0.005 [-0.018, +0.028]
bele B_fragility [W8_W4] - uniform W4 +0.000 [-0.025, +0.025]
bele B_fragility [W8_W4] - mean(2 random) +0.000 [-0.028, +0.028]
bele C_refusalDrop [W8_W4] - uniform W4 -0.005 [-0.030, +0.015]
bele C_refusalDrop [W8_W4] - mean(2 random) -0.005 [-0.033, +0.022]
bele D_composite [W8_W4] - uniform W4 -0.010 [-0.030, +0.010]
bele D_composite [W8_W4] - mean(2 random) -0.010 [-0.037, +0.015]
bele N_negative [W8_W4] - uniform W4 +0.010 [+0.000, +0.025]
bele N_negative [W8_W4] - mean(2 random) +0.010 [-0.013, +0.033]
bele G_gradattr [W8_W4] - uniform W4 +0.000 [-0.015, +0.015]
bele G_gradattr [W8_W4] - mean(2 random) +0.000 [-0.022, +0.022]
refuse_harm E1 uniform W4 - E0 FP16 -0.050 [-0.090, -0.015]
refuse_harm A_importance [FP16_W4] - uniform W4 +0.025 [+0.000, +0.050]
refuse_harm A_importance [FP16_W4] - mean(4 random) +0.016 [-0.007, +0.040]
refuse_harm B_fragility [FP16_W4] - uniform W4 +0.040 [+0.010, +0.070]
refuse_harm B_fragility [FP16_W4] - mean(4 random) +0.031 [+0.005, +0.060]
refuse_harm C_refusalDrop [FP16_W4] - uniform W4 +0.050 [+0.025, +0.080]
refuse_harm C_refusalDrop [FP16_W4] - mean(4 random) +0.041 [+0.014, +0.072]
refuse_harm D_composite [FP16_W4] - uniform W4 +0.045 [+0.020, +0.075]
refuse_harm D_composite [FP16_W4] - mean(4 random) +0.036 [+0.013, +0.062]
refuse_harm N_negative [FP16_W4] - uniform W4 +0.035 [+0.010, +0.060]
refuse_harm N_negative [FP16_W4] - mean(4 random) +0.026 [+0.006, +0.049]
refuse_harm G_gradattr [FP16_W4] - uniform W4 +0.010 [-0.010, +0.030]
refuse_harm G_gradattr [FP16_W4] - mean(4 random) +0.001 [-0.019, +0.021]
refuse_harm A_importance [W8_W4] - uniform W4 +0.025 [+0.000, +0.050]
refuse_harm A_importance [W8_W4] - mean(2 random) +0.022 [+0.003, +0.043]
refuse_harm B_fragility [W8_W4] - uniform W4 +0.040 [+0.010, +0.070]
refuse_harm B_fragility [W8_W4] - mean(2 random) +0.037 [+0.005, +0.072]
refuse_harm C_refusalDrop [W8_W4] - uniform W4 +0.045 [+0.020, +0.075]
refuse_harm C_refusalDrop [W8_W4] - mean(2 random) +0.043 [+0.018, +0.072]
refuse_harm D_composite [W8_W4] - uniform W4 +0.045 [+0.020, +0.075]
refuse_harm D_composite [W8_W4] - mean(2 random) +0.043 [+0.018, +0.070]
refuse_harm N_negative [W8_W4] - uniform W4 +0.030 [+0.010, +0.055]
refuse_harm N_negative [W8_W4] - mean(2 random) +0.028 [+0.005, +0.052]
refuse_harm G_gradattr [W8_W4] - uniform W4 +0.010 [-0.015, +0.035]
refuse_harm G_gradattr [W8_W4] - mean(2 random) +0.007 [-0.015, +0.030]
gap_harm E1 uniform W4 - E0 FP16 -0.170 [-0.295, -0.045]
gap_harm A_importance [FP16_W4] - uniform W4 -0.124 [-0.187, -0.061]
gap_harm A_importance [FP16_W4] - mean(4 random) -0.099 [-0.166, -0.031]
gap_harm B_fragility [FP16_W4] - uniform W4 +0.460 [+0.389, +0.528]
gap_harm B_fragility [FP16_W4] - mean(4 random) +0.486 [+0.422, +0.546]
gap_harm C_refusalDrop [FP16_W4] - uniform W4 +0.706 [+0.635, +0.774]
gap_harm C_refusalDrop [FP16_W4] - mean(4 random) +0.731 [+0.674, +0.785]
gap_harm D_composite [FP16_W4] - uniform W4 +0.202 [+0.148, +0.258]
gap_harm D_composite [FP16_W4] - mean(4 random) +0.228 [+0.173, +0.283]
gap_harm N_negative [FP16_W4] - uniform W4 +0.015 [-0.023, +0.052]
gap_harm N_negative [FP16_W4] - mean(4 random) +0.041 [+0.009, +0.072]
gap_harm G_gradattr [FP16_W4] - uniform W4 -0.022 [-0.063, +0.020]
gap_harm G_gradattr [FP16_W4] - mean(4 random) +0.004 [-0.044, +0.052]
gap_harm A_importance [W8_W4] - uniform W4 -0.181 [-0.251, -0.110]
gap_harm A_importance [W8_W4] - mean(2 random) -0.013 [-0.094, +0.065]
gap_harm B_fragility [W8_W4] - uniform W4 +0.460 [+0.387, +0.529]
gap_harm B_fragility [W8_W4] - mean(2 random) +0.628 [+0.557, +0.699]
gap_harm C_refusalDrop [W8_W4] - uniform W4 +0.623 [+0.557, +0.688]
gap_harm C_refusalDrop [W8_W4] - mean(2 random) +0.791 [+0.731, +0.852]
gap_harm D_composite [W8_W4] - uniform W4 +0.215 [+0.156, +0.276]
gap_harm D_composite [W8_W4] - mean(2 random) +0.383 [+0.315, +0.452]
gap_harm N_negative [W8_W4] - uniform W4 -0.008 [-0.050, +0.031]
gap_harm N_negative [W8_W4] - mean(2 random) +0.160 [+0.115, +0.205]
gap_harm G_gradattr [W8_W4] - uniform W4 -0.013 [-0.057, +0.032]
gap_harm G_gradattr [W8_W4] - mean(2 random) +0.155 [+0.099, +0.213]
refuse_benign E1 uniform W4 - E0 FP16 -0.015 [-0.040, +0.005]
refuse_benign A_importance [FP16_W4] - uniform W4 +0.005 [-0.010, +0.025]
refuse_benign A_importance [FP16_W4] - mean(4 random) +0.005 [-0.007, +0.021]
refuse_benign B_fragility [FP16_W4] - uniform W4 +0.015 [-0.005, +0.040]
refuse_benign B_fragility [FP16_W4] - mean(4 random) +0.015 [-0.004, +0.037]
refuse_benign C_refusalDrop [FP16_W4] - uniform W4 +0.005 [+0.000, +0.015]
refuse_benign C_refusalDrop [FP16_W4] - mean(4 random) +0.005 [-0.003, +0.018]
refuse_benign D_composite [FP16_W4] - uniform W4 +0.005 [-0.010, +0.025]
refuse_benign D_composite [FP16_W4] - mean(4 random) +0.005 [-0.007, +0.021]
refuse_benign N_negative [FP16_W4] - uniform W4 +0.000 [+0.000, +0.000]
refuse_benign N_negative [FP16_W4] - mean(4 random) +0.000 [-0.004, +0.004]
refuse_benign G_gradattr [FP16_W4] - uniform W4 -0.005 [-0.015, +0.000]
refuse_benign G_gradattr [FP16_W4] - mean(4 random) -0.005 [-0.015, +0.000]
refuse_benign A_importance [W8_W4] - uniform W4 +0.005 [-0.010, +0.025]
refuse_benign A_importance [W8_W4] - mean(2 random) +0.007 [-0.005, +0.025]
refuse_benign B_fragility [W8_W4] - uniform W4 +0.015 [-0.005, +0.040]
refuse_benign B_fragility [W8_W4] - mean(2 random) +0.018 [+0.000, +0.040]
refuse_benign C_refusalDrop [W8_W4] - uniform W4 +0.005 [+0.000, +0.015]
refuse_benign C_refusalDrop [W8_W4] - mean(2 random) +0.007 [+0.000, +0.020]
refuse_benign D_composite [W8_W4] - uniform W4 +0.010 [-0.010, +0.030]
refuse_benign D_composite [W8_W4] - mean(2 random) +0.013 [-0.003, +0.030]
refuse_benign N_negative [W8_W4] - uniform W4 +0.000 [+0.000, +0.000]
refuse_benign N_negative [W8_W4] - mean(2 random) +0.003 [+0.000, +0.007]
refuse_benign G_gradattr [W8_W4] - uniform W4 -0.005 [-0.015, +0.000]
refuse_benign G_gradattr [W8_W4] - mean(2 random) -0.003 [-0.007, +0.000]

Limitations: phrase-matcher refusal (no judge model), one model, one quantizer family, random sets may contain fragile layers by chance, CI width reflects the sample sizes above.

Downloads last month
46