# results_avgc_upweight

AVGC-upweighted fine-tune of `stem_ms96` from IRM checkpoints. Multiply AVGC sample loss by **2.5** (other datasets weight 1). Architecture unchanged (ms feats, w96). Init `models/stem_ms96_irm/{held-*,all5}.pt` with causal `-inf` mask restored. EMA 0.995 eval. Smoke AVGC_W∈{2.5,3.0} (40 steps, held-UHD so AVGC is in train) chose **2.5** (frame-acc 0.929 vs 0.924). Full run 800 steps, batch 512. HMM preset p_stay 0.994 scale 0.8. Also scored causal tau-30 high-pass of the evidence (`hp30`; corpus emissions).

When held-out=AVGC, AVGC is absent from training, so the upweight does not apply on that fold (plain ERM continuation from IRM).

Code: `stem_ms96_avgc_upweight.py`. Checkpoints: `models/stem_ms96_avgc_upweight/`. JSON: `results/research/stem_ms96_avgc_upweight.json`. Log: `experiments/stem_ms96_avgc_upweight.log` / `stem_ms96_avgc_upweight_console.log`. Smoke: `experiments/stem_ms96_avgc_upweight_smoke.txt`.

## LODO (preset)

| Held-out | upw H | upw Acc % | IRM H/Acc | SWA H/Acc | stem_ms96 H/Acc | hp30 H | hp30 Acc % |
|---|---:|---:|---|---|---|---:|---:|
| AVGC | 0.8025 | 75.09 | 0.8095/75.12 | 0.8015/74.91 | 0.7945/74.06 | 0.7816 | 70.70 |
| DTU | 0.9006 | 88.70 | 0.8999/88.88 | 0.9003/89.52 | 0.9061/89.85 | 0.8317 | 71.75 |
| KUL | 0.8632 | 81.23 | 0.8702/81.84 | 0.8769/82.35 | 0.8755/82.33 | 0.8371 | 74.40 |
| UHD | 0.8434 | 79.14 | 0.8530/79.96 | 0.8553/80.09 | 0.8407/79.47 | 0.8146 | 73.16 |
| USTC | 0.9065 | 85.43 | 0.9148/85.96 | 0.9161/86.19 | 0.9219/86.36 | 0.8321 | 71.30 |
| **mean** | **0.8632** | **81.92** | **0.8695/82.35** | **0.8700/82.61** | **0.8677/82.42** | | |

## External (preset)

| Set | upw H | upw Acc % | linear_hp30 | IRM | SWA | upw_hp30 H/Acc |
|---|---:|---:|---|---|---|---|
| AASD | 0.6213 | 51.21 | 0.6422/53.12 | 0.6207/51.01 | 0.6202/51.08 | 0.6076/51.08 |
| Carta2026 | 0.5404 | 47.88 | 0.6520/60.54 | 0.5426/48.04 | 0.5507/48.68 | 0.5573/49.52 |

## Targets

| Target | Result |
|---|---|
| AVGC H>=0.85 and Acc>=78 | miss (0.8025/75.09); **worse than IRM** (0.8095/75.12); Acc ≈ IRM |
| UHD H>=0.85 | **lost** (0.8434; was met under IRM 0.8530 / SWA 0.8553) |
| AASD Acc>60 with H>0.63 | miss (0.6213/51.21); Acc flat vs IRM/SWA, below linear_hp30 |
| LODO mean | slightly hurt (0.8632/81.92 vs IRM 0.8695/82.35) |

## Notes

- Smoke preferred milder weight 2.5 over 3.0.
- Held-AVGC fold cannot upweight AVGC samples (LODO); continued IRM fine-tune alone moved AVGC H down.
- Folds with AVGC in train (DTU/KUL/UHD/USTC) did not improve the open targets; UHD H regressed below 0.85.
- No open target flipped to met. AASD Acc unchanged (~51).
