null-edit / mac_4bit /report_codeonly_conf_mac.md
moncefem's picture
The Null Edit: per-token log-probabilities, model replies, per-file metrics, figures (part 6)
8a0878a verified
|
Raw History Blame Contribute Delete
787 Bytes

S line@10; issues from ../results/conf_ids.txt

Model files (issues) all lines @10 code lines only @10 difference (points), p all lines @1 code only @1 hazard on comment/docstring lines (median share) median total hazard (nats)
12B-4bit 69 (58) 41 [29, 52] 57 [45, 68] +15.9 [+7.0, +25.4], p = 0.0008 6 [1, 12] 13 [6, 21] 0.74 64.8

Comment/docstring lines are 30% of non-blank lines (mean over files); every gold line is such a line in 1% of files.

By file length (all lines -> code lines only; n files)

Model under 5,000 tokens 5,000-11,000 tokens over 11,000 tokens
12B-4bit 62 -> 75 (n = 40; share 0.70) 11 -> 33 (n = 27; share 0.77) 0 -> 0 (n = 2; share 0.83)