calib-corpora / metrics.txt
worthant's picture
Rebuild calibration corpus for DeepSeek-V4-Flash-0731 imatrix
bf3d2af verified
Raw
History Blame Contribute Delete
6.66 kB
==============================================================================
calib_train
==============================================================================
documents: 1,087 tokens: 1,868,626
domain docs tokens share med tok p99 tok
------------------------------------------------------------------------------
graphics 409 593,623 31.8% 839 6,855
longctx 18 268,568 14.4% 12,314 39,530
agentic 60 257,837 13.8% 3,905 7,939
code 222 238,365 12.8% 456 6,531
general 31 167,355 9.0% 4,735 15,986
vocab_sweep 65 154,205 8.3% 2,398 2,401
reasoning 174 108,411 5.8% 615 782
structured 108 80,262 4.3% 298 6,928
------------------------------------------------------------------------------
TOTAL 1,087 1,868,626 100.0%
vocabulary coverage (denominator = 129,280 embedding rows):
seen >=1 : 112,574 (87.08%)
seen >=10 : 12,522 ( 9.69%)
seen >=100 : 2,407 ( 1.86%)
document length in tokens: p50=708 p90=4,335 p99=12,251 max=42,081
docs >=8k tokens: 27 (19.8% of all tokens)
deduplication: 5.09% of 12,262 candidates removed (59 exact, 565 near at J>=0.8)
==============================================================================
calib_heldout
==============================================================================
documents: 104 tokens: 141,800
domain docs tokens share med tok p99 tok
------------------------------------------------------------------------------
graphics 39 37,687 26.6% 578 3,009
code 21 31,733 22.4% 423 7,658
vocab_sweep 11 26,377 18.6% 2,398 2,401
general 5 12,690 8.9% 1,173 7,830
agentic 3 12,175 8.6% 3,294 6,916
structured 10 11,867 8.4% 352 5,605
reasoning 15 9,271 6.5% 614 772
------------------------------------------------------------------------------
TOTAL 104 141,800 100.0%
vocabulary coverage (denominator = 129,280 embedding rows):
seen >=1 : 23,139 (17.90%)
seen >=10 : 1,896 ( 1.47%)
seen >=100 : 155 ( 0.12%)
document length in tokens: p50=689 p90=2,523 p99=7,959 max=8,067
docs >=8k tokens: 1 (5.7% of all tokens)
deduplication: 5.09% of 12,262 candidates removed (59 exact, 565 near at J>=0.8)
==============================================================================
eval_neutral
==============================================================================
documents: 66 tokens: 189,407
domain docs tokens share med tok p99 tok
------------------------------------------------------------------------------
general 28 93,738 49.5% 2,107 9,589
eval_code 34 92,016 48.6% 842 17,598
structured 4 3,653 1.9% 844 1,816
------------------------------------------------------------------------------
TOTAL 66 189,407 100.0%
vocabulary coverage (denominator = 129,280 embedding rows):
seen >=1 : 21,748 (16.82%)
seen >=10 : 3,039 ( 2.35%)
seen >=100 : 261 ( 0.20%)
document length in tokens: p50=966 p90=6,830 p99=17,503 max=17,697
docs >=8k tokens: 5 (34.6% of all tokens)
deduplication: 5.09% of 12,262 candidates removed (59 exact, 565 near at J>=0.8)
top-50 tokens in calib_train:
rank id token count share
1 201 '\n' 94,170 5.04%
2 223 ' ' 51,650 2.76%
3 14 ',' 36,580 1.96%
4 16 '.' 24,899 1.33%
5 438 ' =' 23,427 1.25%
6 18 '0' 21,080 1.13%
7 361 ' ' 19,472 1.04%
8 28 ':' 16,596 0.89%
9 19 '1' 14,016 0.75%
10 200 '\t' 13,614 0.73%
11 270 ' the' 13,192 0.71%
12 10 '(' 10,883 0.58%
13 582 ' "' 10,838 0.58%
14 20 '2' 10,011 0.54%
15 982 ' *' 9,613 0.51%
16 528 ' ' 9,535 0.51%
17 11 ')' 8,855 0.47%
18 875 ' {\n' 8,380 0.45%
19 260 ' a' 7,811 0.42%
20 343 ' (' 7,673 0.41%
21 989 ',\n' 7,551 0.40%
22 510 ';\n' 6,864 0.37%
23 17 '/' 6,793 0.36%
24 565 ' -' 6,491 0.35%
25 21 '3' 6,468 0.35%
26 1281 '="' 6,101 0.33%
27 3362 '":' 5,804 0.31%
28 304 ' to' 5,790 0.31%
29 65 '_' 5,788 0.31%
30 1018 '>\n' 5,624 0.30%
31 22 '4' 5,591 0.30%
32 1445 ' ' 5,501 0.29%
33 905 " '" 5,437 0.29%
34 1915 ' //' 4,890 0.26%
35 15 '-' 4,865 0.26%
36 61 '[' 4,591 0.25%
37 2366 '::' 4,584 0.25%
38 271 '\n\n' 4,580 0.25%
39 4 '"' 4,580 0.25%
40 344 ' is' 4,552 0.24%
41 1760 '",' 4,414 0.24%
42 3608 ' `' 4,374 0.23%
43 305 ' and' 4,313 0.23%
44 680 ' {' 4,295 0.23%
45 682 ')\n' 4,182 0.22%
46 4564 '\t\t\t' 4,122 0.22%
47 30 '<' 4,112 0.22%
48 128825 '|DSML|' 4,078 0.22%
49 603 '.\n' 3,944 0.21%
50 1718 '</' 3,943 0.21%