RepoCortex
Plate iiPaused at 4x

Replay the learning pass

200/ 1281

Issues learned in the order Django fixed them. Each one: predict, see the real fix, learn.

The graph changes.
The test set does not.

FIG. 2 An illustrative neighborhood.

Placeholder layout and deltas. Vermilion reweights are illustrative.

Fig. 1 The held-out record

319 chronological fix rows

Checkpoint 183: 41.69%

Held-out File Acc@1 through 183 learned fixesRepoCortex: 41.69%. Frozen graph: 41.07%. Dense vectors: 35.74%. BM25: 34.48%. Baselines are single-point references, not replay curves. The complete data table follows the controls.File Acc@1 / held-out fixes (%)30354045500 fixes: 41.07%92 fixes: 41.69%183 fixes: 41.69%274 fixes: 41.69%366 fixes: 41.38%458 fixes: 41.69%549 fixes: 42.01%640 fixes: 42.32%732 fixes: 42.32%824 fixes: 42.32%915 fixes: 42.63%1006 fixes: 42.63%1098 fixes: 44.2%1190 fixes: 45.14%1281 fixes: 44.51%RepoCortex41.69%Frozen graph41.07%Dense vectors35.74%BM2534.48%018336654973291510981281Merged fixes learned, in time order

Lines join measured checkpoints, including the dip. Between checkpoints, the score stays at the last measurement. Function Acc@10 ends at 48.28%. Baselines are single-point references.

FIG. 3 Checkpoint register / File Acc@1, not a mutation log.

FIG. 4 Expanding window on the training stream, not the holdout. Each of 1281 issues is scored before the graph learns from it.

Learning, Acc@10
47.23%
Frozen, Acc@10
43.87%
Gap
+3.4 pp
Last 100, learning / frozen
46% / 41%
03206419611281
Read the checkpoint data
File Acc@1, same 319 held-out fix rows at every checkpoint
Fixes learnedRepoCortexDenseFrozenBM25
041.07%35.74%41.07%34.48%
9241.69%35.74%41.07%34.48%
18341.69%35.74%41.07%34.48%
27441.69%35.74%41.07%34.48%
36641.38%35.74%41.07%34.48%
45841.69%35.74%41.07%34.48%
54942.01%35.74%41.07%34.48%
64042.32%35.74%41.07%34.48%
73242.32%35.74%41.07%34.48%
82442.32%35.74%41.07%34.48%
91542.63%35.74%41.07%34.48%
100642.63%35.74%41.07%34.48%
109844.2%35.74%41.07%34.48%
119045.14%35.74%41.07%34.48%
128144.51%35.74%41.07%34.48%

The frozen reference is this pass's step 0 graph, with learning off. Playback starts paused, including with reduced motion.