File size: 2,599 Bytes
5386334 f8dff9a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 | # 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).**
1. **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.
2. **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).
3. **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.py` in-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.
|