exp011a_fused_multiband — the multiband mechanism on REAL DATA
(+ the segmentation blob-target cache) — CANDIDATE s0; s1 running
Questions. (1) Does real fused data (multi-entity photos, natural captions) change the band economics that uniform synthetic data couldn't pay for? (2) Build the segmentation blob targets for the HIGH-band blobbing role (loss design held for review — new mechanism, pacing law).
Design. qwen-deepfashion-fused, 4352 rows (dataset audit gates enforced; all rows passed), caption_joycaption at nl77, same certified beds as exp008 (multiband3 vs mono48, uniform eps objective, 3000 steps, pure Adam wd=0).
Results (results.json):
| arm | overall val | band0 (LOW) | band1 (MID) | band2 (HIGH) |
|---|---|---|---|---|
| frozen | 0.142435 | 0.293931 | 0.071764 | 0.010422 |
| multiband3 | 0.135215 (−5.1%) | 0.280633 | 0.066728 | 0.009562 |
| monolith r48 | 0.134184 (−5.8%) | 0.278893 | 0.065772 | 0.009564 |
Lesions: 3/3 surgical, with LARGER asymmetries than synthetic (band0 own 0.0130 vs 0.000062 cross — 200×).
Findings (candidate, s0).
- Real data pays adapters 2–3× more than synthetic (−5.1/−5.8% vs −2.2/−2.4% in exp008) — the richer distribution has more for small adapters to learn.
- The structural story replicates on real data: surgical band specialists, monolith edge persists under a uniform objective, HIGH-band own-band win reappears at a hair (0.009562 vs 0.009564).
- Blob targets: built, verified, cached — after a schema lesson: the
first rasterizer pass returned 100% EMPTY (fused_json polygons are FLAT
[x1,y1,...] lists in a declared
coord_space, not the assumed pair-list NORM_0_1000; the emptiness counter caught it — silent-zero discipline). Corrected rasterizer: 100% non-empty, mean occupancy 0.293, coord_space norm_0_1000 confirmed on every row.fix_blob_targets.pyin-package.
Next. The blob LOSS (how coarse occupancy couples to the HIGH band's x0-estimate) goes to the 8am design review with targets ready. s1 replication of this bed lands ~07:30Z.
Caveats. s0; 4352 rows single dataset; blob maps unused in this training (uniform objective by design). Cost ≈ 2.9 GPU-h incl. cache.
SEED-1 REPLICATION (2026-07-18, results_s1.json) — 2-SEED ON REAL DATA
mb3 0.135245/0.135215, mono 0.134221/0.134184 — the 2-3x real-data payoff, the monolith edge, and 3/3 surgical lesions (band0 own 0.01292/0.01296 vs cross 0.00007) all replicate near-exactly. The fused-data multiband story is seed-robust.