null-edit / mac_4bit /report_edit_mac_lenient.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
4.86 kB

98 files, 83 issues; one greedy search/replace edit per file and arm; oracle file; nothing executed; a reply that hit the token cap is scored if its SEARCH block is complete (sensitivity analysis)

Arm valid ambiguous truncated noop on_gold near_gold gold_unit adds_gold on_hint on_gold if valid median tokens
no hint 83 [75, 90] 2 [0, 5] 29 [19, 39] 11 [5, 18] 32 [22, 42] 40 [29, 50] 47 [36, 57] 5 [1, 11] 0 [0, 0] 38 [27, 49] 336
hint: top-5 null-edit lines 84 [76, 91] 1 [0, 3] 27 [18, 35] 8 [3, 14] 31 [20, 42] 43 [32, 54] 49 [38, 60] 6 [1, 12] 58 [48, 69] 37 [25, 48] 305
hint: top-5 wrong-issue lines (placebo) 76 [67, 84] 0 [0, 0] 33 [23, 42] 16 [9, 24] 21 [13, 31] 24 [16, 34] 32 [22, 42] 3 [0, 8] 18 [11, 26] 28 [18, 40] 352
hint: gold lines (ceiling) 91 [85, 96] 1 [0, 3] 18 [11, 27] 13 [7, 20] 47 [36, 58] 62 [52, 72] 66 [56, 76] 6 [1, 12] 44 [33, 55] 52 [40, 62] 262

Primary contrasts (on_gold, percentage points, instance-clustered bootstrap, unadjusted p):

  • hint: top-5 null-edit lines − hint: top-5 wrong-issue lines (placebo): +9.2 [-1.0, +19.1], p = 0.0836
  • hint: top-5 null-edit lines − no hint: -1.0 [-9.6, +7.7], p = 0.9160

Secondary contrasts (Holm-adjusted over this list):

  • near_gold: hint: top-5 null-edit lines − no hint: +3.1 [-5.4, +12.1], p = 0.5564, Holm 1.000
  • gold_unit: hint: top-5 null-edit lines − no hint: +2.0 [-7.4, +12.1], p = 0.7564, Holm 1.000
  • adds_gold: hint: top-5 null-edit lines − no hint: +1.0 [-2.1, +4.8], p = 0.7864, Holm 1.000
  • valid: hint: top-5 null-edit lines − no hint: +1.0 [-8.1, +10.4], p = 0.9184, Holm 1.000
  • near_gold: hint: top-5 null-edit lines − hint: top-5 wrong-issue lines (placebo): +18.4 [+8.4, +28.6], p = 0.0010, Holm 0.021
  • gold_unit: hint: top-5 null-edit lines − hint: top-5 wrong-issue lines (placebo): +17.3 [+7.2, +27.5], p = 0.0012, Holm 0.024
  • adds_gold: hint: top-5 null-edit lines − hint: top-5 wrong-issue lines (placebo): +3.1 [+0.0, +6.9], p = 0.0980, Holm 1.000
  • valid: hint: top-5 null-edit lines − hint: top-5 wrong-issue lines (placebo): +8.2 [-3.1, +19.8], p = 0.1878, Holm 1.000
  • on_gold: hint: top-5 wrong-issue lines (placebo) − no hint: -10.2 [-19.8, +0.0], p = 0.0654, Holm 0.850
  • near_gold: hint: top-5 wrong-issue lines (placebo) − no hint: -15.3 [-25.2, -5.2], p = 0.0068, Holm 0.095
  • gold_unit: hint: top-5 wrong-issue lines (placebo) − no hint: -15.3 [-25.0, -5.1], p = 0.0050, Holm 0.085
  • adds_gold: hint: top-5 wrong-issue lines (placebo) − no hint: -2.0 [-5.1, +0.0], p = 0.2682, Holm 1.000
  • valid: hint: top-5 wrong-issue lines (placebo) − no hint: -7.1 [-17.7, +2.8], p = 0.1920, Holm 1.000
  • on_gold: hint: gold lines (ceiling) − hint: top-5 null-edit lines: +16.3 [+5.2, +27.6], p = 0.0054, Holm 0.086
  • near_gold: hint: gold lines (ceiling) − hint: top-5 null-edit lines: +19.4 [+8.8, +29.9], p = 0.0004, Holm 0.009
  • gold_unit: hint: gold lines (ceiling) − hint: top-5 null-edit lines: +17.3 [+5.5, +28.7], p = 0.0040, Holm 0.072
  • adds_gold: hint: gold lines (ceiling) − hint: top-5 null-edit lines: +0.0 [-3.1, +3.0], p = 1.0000, Holm 1.000
  • valid: hint: gold lines (ceiling) − hint: top-5 null-edit lines: +7.1 [-1.1, +15.1], p = 0.1106, Holm 1.000
  • on_gold: hint: gold lines (ceiling) − no hint: +15.3 [+4.7, +26.5], p = 0.0056, Holm 0.086
  • near_gold: hint: gold lines (ceiling) − no hint: +22.4 [+11.6, +34.0], p = 0.0000, Holm 0.000
  • gold_unit: hint: gold lines (ceiling) − no hint: +19.4 [+7.3, +31.9], p = 0.0014, Holm 0.027
  • adds_gold: hint: gold lines (ceiling) − no hint: +1.0 [-2.1, +4.9], p = 0.7808, Holm 1.000
  • valid: hint: gold lines (ceiling) − no hint: +8.2 [+0.0, +17.0], p = 0.0824, Holm 0.989

Does the model follow the hint, or use it? (H arm)

  • hint contains a gold line (47 files): on_gold with hint 60 [44, 75], without hint 53 [37, 68]; edit touches a hinted line 72 [58, 85]
  • hint contains no gold line (51 files): on_gold with hint 4 [0, 10], without hint 12 [4, 20]; edit touches a hinted line 45 [31, 60]
  • hint: top-5 null-edit lines: hint contains a gold line in 48% of files; edit touches a hinted line in 58%
  • hint: top-5 wrong-issue lines (placebo): hint contains a gold line in 7% of files; edit touches a hinted line in 18%
  • hint: gold lines (ceiling): hint contains a gold line in 100% of files; edit touches a hinted line in 44%

On the same files the map alone: top-1 line is gold 21%, a gold line in the top 5 48%.

Files the greedy rewrite would leave unchanged (8 files): on_gold no hint 38 [12, 71]; hint: top-5 null-edit lines 0 [0, 0]; hint: top-5 wrong-issue lines (placebo) 12 [0, 43]; hint: gold lines (ceiling) 50 [22, 86]