Spaces:
Running
Running
| { | |
| "_meta": { | |
| "note": "single mode: best measured KL that sits fully on the hardware wins; no band caps (if Q8_0 fits, it wins). emergency only when nothing else fits. avoid is never dropped outright: dominance + kl sort handle repo-dependent availability.", | |
| "bands": { | |
| "archival": "near-lossless, take when headroom allows (quality only)", | |
| "avoid": "dominated in size+quality by another quant; only wins when its dominator is absent from the repo", | |
| "default": "rows 5-8 of PRIORITY.md; balanced picks the best-kl fitting one of these", | |
| "emergency": "below the quality knee; picked only when nothing else fits, with a warning", | |
| "knee": "min viable (UD-IQ3_XXS, ~4.4bpw)", | |
| "step": "step up from default when quality is the priority", | |
| "tight": "tight budget, between knee and default" | |
| }, | |
| "method": "llama-perplexity --kl-divergence vs family F16/BF16 ref, 6x repeated calibration text (~2500 tokens, ctx 1024, full GPU offload, stock llama.cpp, RTX 5070). norm_kl: 0 = full-precision ref, 1 = family worst. pooled over LFM2.5-1.2B + Qwen3.5-0.8B + gemma-3-1b-it; ranking was consistent across all 3 families. quality = round(100*(1-log10(1+1000*norm_kl)/3)).", | |
| "quality_floor_bpw": 4.5, | |
| "rule_of_thumb": "nothing under ~4.5bpw. default to the default band. avoid band is never the right answer when its dominator is in the same repo.", | |
| "source": "quant-bench: 150 GGUFs, 12 repos, results.csv + PRIORITY.md (2026-10-14)" | |
| }, | |
| "quants": [ | |
| { | |
| "band": "knee", | |
| "bpw": 4.4, | |
| "n": 2, | |
| "norm_kl": 0.1586, | |
| "q": 27, | |
| "quant": "UD-IQ3_XXS" | |
| }, | |
| { | |
| "band": "tight", | |
| "bpw": 4.98, | |
| "n": 3, | |
| "norm_kl": 0.1166, | |
| "q": 31, | |
| "quant": "UD-Q3_K_XL" | |
| }, | |
| { | |
| "band": "tight", | |
| "bpw": 5.46, | |
| "n": 4, | |
| "norm_kl": 0.1011, | |
| "q": 33, | |
| "quant": "IQ4_XS" | |
| }, | |
| { | |
| "band": "tight", | |
| "bpw": 5.56, | |
| "n": 4, | |
| "norm_kl": 0.0811, | |
| "q": 36, | |
| "quant": "IQ4_NL" | |
| }, | |
| { | |
| "band": "default", | |
| "bpw": 5.72, | |
| "n": 3, | |
| "norm_kl": 0.0541, | |
| "q": 42, | |
| "quant": "UD-Q4_K_XL" | |
| }, | |
| { | |
| "band": "default", | |
| "bpw": 5.81, | |
| "n": 11, | |
| "norm_kl": 0.0465, | |
| "q": 44, | |
| "quant": "Q4_K_M" | |
| }, | |
| { | |
| "band": "default", | |
| "bpw": 6.03, | |
| "n": 2, | |
| "norm_kl": 0.0245, | |
| "q": 53, | |
| "quant": "Q3_K_XL" | |
| }, | |
| { | |
| "band": "default", | |
| "bpw": 6.32, | |
| "n": 3, | |
| "norm_kl": 0.0188, | |
| "q": 57, | |
| "quant": "UD-Q5_K_XL" | |
| }, | |
| { | |
| "band": "step", | |
| "bpw": 7.03, | |
| "n": 2, | |
| "norm_kl": 0.0118, | |
| "q": 63, | |
| "quant": "Q5_K_L" | |
| }, | |
| { | |
| "band": "step", | |
| "bpw": 7.21, | |
| "n": 10, | |
| "norm_kl": 0.0044, | |
| "q": 76, | |
| "quant": "Q6_K" | |
| }, | |
| { | |
| "band": "step", | |
| "bpw": 7.65, | |
| "n": 3, | |
| "norm_kl": 0.0032, | |
| "q": 79, | |
| "quant": "UD-Q6_K_XL" | |
| }, | |
| { | |
| "band": "step", | |
| "bpw": 7.9, | |
| "n": 2, | |
| "norm_kl": 0.0026, | |
| "q": 81, | |
| "quant": "Q6_K_L" | |
| }, | |
| { | |
| "band": "archival", | |
| "bpw": 8.5, | |
| "n": 12, | |
| "norm_kl": 0.0012, | |
| "q": 89, | |
| "quant": "Q8_0" | |
| }, | |
| { | |
| "band": "archival", | |
| "bpw": 11.11, | |
| "n": 3, | |
| "norm_kl": 0.0009, | |
| "q": 91, | |
| "quant": "UD-Q8_K_XL" | |
| }, | |
| { | |
| "band": "emergency", | |
| "bpw": 4.21, | |
| "n": 2, | |
| "norm_kl": 0.2229, | |
| "q": 22, | |
| "quant": "UD-IQ2_M" | |
| }, | |
| { | |
| "band": "emergency", | |
| "bpw": 4.72, | |
| "n": 2, | |
| "norm_kl": 0.3222, | |
| "q": 16, | |
| "quant": "IQ2_M" | |
| }, | |
| { | |
| "band": "emergency", | |
| "bpw": 4.91, | |
| "n": 5, | |
| "norm_kl": 0.4585, | |
| "q": 11, | |
| "quant": "Q2_K" | |
| }, | |
| { | |
| "band": "emergency", | |
| "bpw": 4.93, | |
| "n": 4, | |
| "norm_kl": 0.4761, | |
| "q": 11, | |
| "quant": "Q2_K_L" | |
| }, | |
| { | |
| "band": "emergency", | |
| "bpw": 4.46, | |
| "n": 1, | |
| "norm_kl": 0.5426, | |
| "q": 9, | |
| "quant": "UD-IQ1_M" | |
| }, | |
| { | |
| "band": "emergency", | |
| "bpw": 4.44, | |
| "n": 1, | |
| "norm_kl": 0.7422, | |
| "q": 4, | |
| "quant": "UD-IQ1_S" | |
| }, | |
| { | |
| "band": "emergency", | |
| "bpw": 3.99, | |
| "n": 2, | |
| "norm_kl": 1.0, | |
| "q": 0, | |
| "quant": "UD-IQ2_XXS" | |
| }, | |
| { | |
| "band": "avoid", | |
| "bpw": 6.37, | |
| "n": 7, | |
| "norm_kl": 0.0259, | |
| "q": 52, | |
| "quant": "Q5_K_M" | |
| }, | |
| { | |
| "band": "avoid", | |
| "bpw": 6.29, | |
| "n": 6, | |
| "norm_kl": 0.0368, | |
| "q": 47, | |
| "quant": "Q5_K_S" | |
| }, | |
| { | |
| "band": "avoid", | |
| "bpw": 6.51, | |
| "n": 2, | |
| "norm_kl": 0.0401, | |
| "q": 46, | |
| "quant": "Q4_K_L" | |
| }, | |
| { | |
| "band": "avoid", | |
| "bpw": 5.73, | |
| "n": 6, | |
| "norm_kl": 0.0596, | |
| "q": 41, | |
| "quant": "Q4_K_S" | |
| }, | |
| { | |
| "band": "avoid", | |
| "bpw": 5.85, | |
| "n": 4, | |
| "norm_kl": 0.0796, | |
| "q": 36, | |
| "quant": "Q3_K_L" | |
| }, | |
| { | |
| "band": "avoid", | |
| "bpw": 5.74, | |
| "n": 5, | |
| "norm_kl": 0.1154, | |
| "q": 31, | |
| "quant": "Q4_1" | |
| }, | |
| { | |
| "band": "avoid", | |
| "bpw": 5.39, | |
| "n": 2, | |
| "norm_kl": 0.12, | |
| "q": 31, | |
| "quant": "IQ3_M" | |
| }, | |
| { | |
| "band": "avoid", | |
| "bpw": 5.05, | |
| "n": 2, | |
| "norm_kl": 0.1219, | |
| "q": 30, | |
| "quant": "IQ3_XXS" | |
| }, | |
| { | |
| "band": "avoid", | |
| "bpw": 5.25, | |
| "n": 6, | |
| "norm_kl": 0.1243, | |
| "q": 30, | |
| "quant": "Q3_K_M" | |
| }, | |
| { | |
| "band": "avoid", | |
| "bpw": 5.31, | |
| "n": 2, | |
| "norm_kl": 0.1557, | |
| "q": 27, | |
| "quant": "IQ3_XS" | |
| }, | |
| { | |
| "band": "avoid", | |
| "bpw": 5.28, | |
| "n": 8, | |
| "norm_kl": 0.2075, | |
| "q": 23, | |
| "quant": "Q4_0" | |
| }, | |
| { | |
| "band": "avoid", | |
| "bpw": 5.0, | |
| "n": 6, | |
| "norm_kl": 0.2849, | |
| "q": 18, | |
| "quant": "Q3_K_S" | |
| }, | |
| { | |
| "band": "avoid", | |
| "bpw": 4.37, | |
| "n": 3, | |
| "norm_kl": 0.5345, | |
| "q": 9, | |
| "quant": "UD-Q2_K_XL" | |
| } | |
| ] | |
| } |