# results_avgc_curriculum

AVGC-hard **curriculum** fine-tune of `stem_ms96` from IRM (≠ `avgc-upweight` loss λ). Frozen **IRM↔synth teacher** (w=0.5 logit blend) scores every train frame; hardness = BCE (smoke pick) or 1/(|logit|+ε). After 20% warm-up: ramp hardness-weighted window sampling to `HARD_MIX_MAX=0.75`, and when AVGC∈train ramp AVGC batch fraction to `MAX_AVGC_FRAC=0.45`. Student init `models/stem_ms96_irm/{held-*,all5}.pt`; EMA 0.995; HMM 0.994/0.8 + hp30. Never AASD/Carta labels.

Smoke HARD_KIND∈{bce,abs_logit} (40 steps, held-UHD) → **bce** (frame 0.929 vs 0.927). Full 1000 steps, batch 512.

Code: `stem_ms96_avgc_curriculum.py`. Checkpoints: `models/stem_ms96_avgc_curriculum/`. JSON: `results/research/stem_ms96_avgc_curriculum.json`. Log: `experiments/stem_ms96_avgc_curriculum.log` / `_console.log`.

## LODO (preset)

| Held-out | Curr H | Curr Acc % | IRM | SWA | IRM↔synth fuse | hp30 H | hp30 Acc % |
|---|---:|---:|---|---|---|---:|---:|
| AVGC | **0.7499** | **72.37** | 0.8095/75.12 | 0.8015/74.91 | **0.8192/75.52** | 0.7033 | 66.87 |
| DTU | 0.8833 | 87.48 | 0.8999/88.88 | 0.9003/89.52 | — | 0.7980 | 66.76 |
| KUL | 0.8211 | 77.52 | 0.8702/81.84 | 0.8769/82.35 | — | 0.7784 | 67.98 |
| UHD | 0.8015 | 76.51 | 0.8530/79.96 | 0.8553/80.09 | 0.8614 | 0.7442 | 68.34 |
| USTC | 0.8661 | 82.11 | 0.9148/85.96 | 0.9161/86.19 | — | 0.8026 | 67.13 |
| **mean** | **0.8244** | **79.20** | **0.8695/82.35** | **0.8700/82.61** | **0.8790/83.11** | | |

## External (preset)

| Set | Curr H | Curr Acc % | linear_hp30 | SW2/swmix | Curr_hp30 H/Acc |
|---|---:|---:|---|---|---|
| AASD | 0.6194 | 50.72 | 0.6422/53.12 | **53.53** | 0.6081/51.08 |
| Carta2026 | 0.5323 | 47.23 | 0.6520/60.54 | 59.71 | 0.5525/48.63 |

## Targets

| Target | Result |
|---|---|
| AVGC H≥0.85 and Acc≥78 | **miss** (0.7499/72.37); **Δ−0.0693 H / −3.15 Acc vs fuse tip 0.8192/75.52**; worse than IRM/SWA/upweight |
| UHD H≥0.85 | **lost** (0.8015; was met under IRM/SWA/fuse) |
| AASD Acc>60 with H>0.63 | miss (0.6194/50.72); Acc flat ~51 |
| LODO mean | hurt (0.8244/79.20 vs IRM 0.8695/82.35) |

## Notes

- Smoke preferred BCE hardness over abs-logit.
- Held-AVGC has no AVGC in train → curriculum = hard-window upsample on DTU/KUL/UHD/USTC only; **destroys AVGC transfer** (same qualitative failure family as avgc-upweight, worse magnitude).
- Folds with AVGC∈train also regress LODO H/Acc vs IRM; UHD drops below 0.85.
- No open target moved. Prefer keep IRM↔synth fuse AVGC tip; do not reopen AVGC loss-λ or hard-window curricula on stem_ms96.

## NEXT IDEA

**Confidence-gated residual head on frozen IRM↔synth fuse**: freeze fuse stream; train a tiny residual logit head only on low-|fuse| frames (teacher-hard), emit `fuse + α·resid·gate(|fuse|)` with LODO-selected α/gate. Distinct from reweighting/mining the full stem (both upweight + curriculum failed).
