Hugging Face's logo

Spaces:
spl16
/
local-model-explorer
Running

App Files Community
spl16 commited on
Commit
13b32b1
·
verified ·
1 Parent(s): 09ba564

Upload quant_scores.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. quant_scores.json +292 -1
quant_scores.json CHANGED
@@ -1 +1,292 @@
1
- {"_meta":{"note":"band order: balanced/small cap at default; quality goes to archival. 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"}]}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_meta": {
3
+ "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.",
4
+ "bands": {
5
+ "archival": "near-lossless, take when headroom allows (quality only)",
6
+ "avoid": "dominated in size+quality by another quant; only wins when its dominator is absent from the repo",
7
+ "default": "rows 5-8 of PRIORITY.md; balanced picks the best-kl fitting one of these",
8
+ "emergency": "below the quality knee; picked only when nothing else fits, with a warning",
9
+ "knee": "min viable (UD-IQ3_XXS, ~4.4bpw)",
10
+ "step": "step up from default when quality is the priority",
11
+ "tight": "tight budget, between knee and default"
12
+ },
13
+ "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)).",
14
+ "quality_floor_bpw": 4.5,
15
+ "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.",
16
+ "source": "quant-bench: 150 GGUFs, 12 repos, results.csv + PRIORITY.md (2026-10-14)"
17
+ },
18
+ "quants": [
19
+ {
20
+ "band": "knee",
21
+ "bpw": 4.4,
22
+ "n": 2,
23
+ "norm_kl": 0.1586,
24
+ "q": 27,
25
+ "quant": "UD-IQ3_XXS"
26
+ },
27
+ {
28
+ "band": "tight",
29
+ "bpw": 4.98,
30
+ "n": 3,
31
+ "norm_kl": 0.1166,
32
+ "q": 31,
33
+ "quant": "UD-Q3_K_XL"
34
+ },
35
+ {
36
+ "band": "tight",
37
+ "bpw": 5.46,
38
+ "n": 4,
39
+ "norm_kl": 0.1011,
40
+ "q": 33,
41
+ "quant": "IQ4_XS"
42
+ },
43
+ {
44
+ "band": "tight",
45
+ "bpw": 5.56,
46
+ "n": 4,
47
+ "norm_kl": 0.0811,
48
+ "q": 36,
49
+ "quant": "IQ4_NL"
50
+ },
51
+ {
52
+ "band": "default",
53
+ "bpw": 5.72,
54
+ "n": 3,
55
+ "norm_kl": 0.0541,
56
+ "q": 42,
57
+ "quant": "UD-Q4_K_XL"
58
+ },
59
+ {
60
+ "band": "default",
61
+ "bpw": 5.81,
62
+ "n": 11,
63
+ "norm_kl": 0.0465,
64
+ "q": 44,
65
+ "quant": "Q4_K_M"
66
+ },
67
+ {
68
+ "band": "default",
69
+ "bpw": 6.03,
70
+ "n": 2,
71
+ "norm_kl": 0.0245,
72
+ "q": 53,
73
+ "quant": "Q3_K_XL"
74
+ },
75
+ {
76
+ "band": "default",
77
+ "bpw": 6.32,
78
+ "n": 3,
79
+ "norm_kl": 0.0188,
80
+ "q": 57,
81
+ "quant": "UD-Q5_K_XL"
82
+ },
83
+ {
84
+ "band": "step",
85
+ "bpw": 7.03,
86
+ "n": 2,
87
+ "norm_kl": 0.0118,
88
+ "q": 63,
89
+ "quant": "Q5_K_L"
90
+ },
91
+ {
92
+ "band": "step",
93
+ "bpw": 7.21,
94
+ "n": 10,
95
+ "norm_kl": 0.0044,
96
+ "q": 76,
97
+ "quant": "Q6_K"
98
+ },
99
+ {
100
+ "band": "step",
101
+ "bpw": 7.65,
102
+ "n": 3,
103
+ "norm_kl": 0.0032,
104
+ "q": 79,
105
+ "quant": "UD-Q6_K_XL"
106
+ },
107
+ {
108
+ "band": "step",
109
+ "bpw": 7.9,
110
+ "n": 2,
111
+ "norm_kl": 0.0026,
112
+ "q": 81,
113
+ "quant": "Q6_K_L"
114
+ },
115
+ {
116
+ "band": "archival",
117
+ "bpw": 8.5,
118
+ "n": 12,
119
+ "norm_kl": 0.0012,
120
+ "q": 89,
121
+ "quant": "Q8_0"
122
+ },
123
+ {
124
+ "band": "archival",
125
+ "bpw": 11.11,
126
+ "n": 3,
127
+ "norm_kl": 0.0009,
128
+ "q": 91,
129
+ "quant": "UD-Q8_K_XL"
130
+ },
131
+ {
132
+ "band": "emergency",
133
+ "bpw": 4.21,
134
+ "n": 2,
135
+ "norm_kl": 0.2229,
136
+ "q": 22,
137
+ "quant": "UD-IQ2_M"
138
+ },
139
+ {
140
+ "band": "emergency",
141
+ "bpw": 4.72,
142
+ "n": 2,
143
+ "norm_kl": 0.3222,
144
+ "q": 16,
145
+ "quant": "IQ2_M"
146
+ },
147
+ {
148
+ "band": "emergency",
149
+ "bpw": 4.91,
150
+ "n": 5,
151
+ "norm_kl": 0.4585,
152
+ "q": 11,
153
+ "quant": "Q2_K"
154
+ },
155
+ {
156
+ "band": "emergency",
157
+ "bpw": 4.93,
158
+ "n": 4,
159
+ "norm_kl": 0.4761,
160
+ "q": 11,
161
+ "quant": "Q2_K_L"
162
+ },
163
+ {
164
+ "band": "emergency",
165
+ "bpw": 4.46,
166
+ "n": 1,
167
+ "norm_kl": 0.5426,
168
+ "q": 9,
169
+ "quant": "UD-IQ1_M"
170
+ },
171
+ {
172
+ "band": "emergency",
173
+ "bpw": 4.44,
174
+ "n": 1,
175
+ "norm_kl": 0.7422,
176
+ "q": 4,
177
+ "quant": "UD-IQ1_S"
178
+ },
179
+ {
180
+ "band": "emergency",
181
+ "bpw": 3.99,
182
+ "n": 2,
183
+ "norm_kl": 1.0,
184
+ "q": 0,
185
+ "quant": "UD-IQ2_XXS"
186
+ },
187
+ {
188
+ "band": "avoid",
189
+ "bpw": 6.37,
190
+ "n": 7,
191
+ "norm_kl": 0.0259,
192
+ "q": 52,
193
+ "quant": "Q5_K_M"
194
+ },
195
+ {
196
+ "band": "avoid",
197
+ "bpw": 6.29,
198
+ "n": 6,
199
+ "norm_kl": 0.0368,
200
+ "q": 47,
201
+ "quant": "Q5_K_S"
202
+ },
203
+ {
204
+ "band": "avoid",
205
+ "bpw": 6.51,
206
+ "n": 2,
207
+ "norm_kl": 0.0401,
208
+ "q": 46,
209
+ "quant": "Q4_K_L"
210
+ },
211
+ {
212
+ "band": "avoid",
213
+ "bpw": 5.73,
214
+ "n": 6,
215
+ "norm_kl": 0.0596,
216
+ "q": 41,
217
+ "quant": "Q4_K_S"
218
+ },
219
+ {
220
+ "band": "avoid",
221
+ "bpw": 5.85,
222
+ "n": 4,
223
+ "norm_kl": 0.0796,
224
+ "q": 36,
225
+ "quant": "Q3_K_L"
226
+ },
227
+ {
228
+ "band": "avoid",
229
+ "bpw": 5.74,
230
+ "n": 5,
231
+ "norm_kl": 0.1154,
232
+ "q": 31,
233
+ "quant": "Q4_1"
234
+ },
235
+ {
236
+ "band": "avoid",
237
+ "bpw": 5.39,
238
+ "n": 2,
239
+ "norm_kl": 0.12,
240
+ "q": 31,
241
+ "quant": "IQ3_M"
242
+ },
243
+ {
244
+ "band": "avoid",
245
+ "bpw": 5.05,
246
+ "n": 2,
247
+ "norm_kl": 0.1219,
248
+ "q": 30,
249
+ "quant": "IQ3_XXS"
250
+ },
251
+ {
252
+ "band": "avoid",
253
+ "bpw": 5.25,
254
+ "n": 6,
255
+ "norm_kl": 0.1243,
256
+ "q": 30,
257
+ "quant": "Q3_K_M"
258
+ },
259
+ {
260
+ "band": "avoid",
261
+ "bpw": 5.31,
262
+ "n": 2,
263
+ "norm_kl": 0.1557,
264
+ "q": 27,
265
+ "quant": "IQ3_XS"
266
+ },
267
+ {
268
+ "band": "avoid",
269
+ "bpw": 5.28,
270
+ "n": 8,
271
+ "norm_kl": 0.2075,
272
+ "q": 23,
273
+ "quant": "Q4_0"
274
+ },
275
+ {
276
+ "band": "avoid",
277
+ "bpw": 5.0,
278
+ "n": 6,
279
+ "norm_kl": 0.2849,
280
+ "q": 18,
281
+ "quant": "Q3_K_S"
282
+ },
283
+ {
284
+ "band": "avoid",
285
+ "bpw": 4.37,
286
+ "n": 3,
287
+ "norm_kl": 0.5345,
288
+ "q": 9,
289
+ "quant": "UD-Q2_K_XL"
290
+ }
291
+ ]
292
+ }