OPD / TCOD ALFWorld checkpoints (B200 multiseed campaign)
Checkpoints from the on-policy distillation (OPD) baseline reproduction and turn-level curriculum (TCOD / entropy-adaptive) experiments on ALFWorld, student = Qwen2.5-3B-Instruct, teacher = GiGPO-Qwen2.5-7B-Instruct-ALFWorld. Trained on B200 (machines b200-158 and b200-203), 250 steps (extension arms to 310), evaluated with the frozen full-274-task protocol (137 valid_seen + 137 valid_unseen, greedy).
Code, configs, logs and analysis: https://github.com/Sisyphe-lee/opd-baseline-repro
Endpoint results (multiseed campaign, 2026-08-16 → 08-19)
| Method | n seeds | mean±std (tasks solved / 274) | min |
|---|---|---|---|
| adaptive τ=0.100 | 8 | 225.8 ± 9.8 | 206 |
| TCOD-F2B | 7 | 222.9 ± 5.9 | 213 |
| Vanilla OPD | 6 | 206.2 ± 19.0 | 173 |
- adaptive vs vanilla: Δ+19.6, Mann-Whitney p=0.020; TCOD vs vanilla: Δ+16.7, p=0.045.
- adaptive ≈ TCOD at endpoint; adaptive needs no η/k_start tuning and trains only ~70–80% of turns.
- Checkpoint soup (equal average of steps 220/240/250) reduces endpoint variance: σ 11.0 → 6.3, worst seed 206 → 216.
- Extending 250→310 mildly helps TCOD (+3~4) and does not help adaptive.
See analysis/multiseed_curves_20260817/ in the GitHub repo for the full
campaign summary, extension-310 study, and seed45 failure analysis.
Repository layout
62 checkpoints, ~421 GB. Every leaf directory <method>/<seed>/<step-or-soup>
is a complete HF-format model (2 safetensors shards + tokenizer), loadable with
AutoModelForCausalLM.from_pretrained("Her77/opd-baseline-repro-checkpoints", subfolder="<leaf>").
adaptive_t0100 (entropy-adaptive curriculum, τ=0.100)
| seed | machine | contents | endpoint full274 |
|---|---|---|---|
| seed42 | b200-203 | step220, step240, step250, step270, step290, step310, soup220_250 | 230 (soup 218) |
| seed43 | b200-203 | step220, step240, step250, step270, step290, step310, soup220_250 | 234 (soup 232) |
| seed45 | b200-158 | step220, step240, step250, soup220_250 | 206 (weak seed; soup 216) |
| seed46 | b200-158 | step220, step240, step250, soup220_250 | 221 |
| seed47 | b200-158 | step220, step240, step250, soup220_250 | 228 |
tcod_f2b (TCOD front-to-back curriculum)
| seed | contents | endpoint full274 |
|---|---|---|
| seed42 | step20…step240 every 20 steps + step250 (frozen-config b158 repro; per-step sweep for training curves) | 223 |
| seed43 | step250, step270, step290, step310 | 224 (310: 228) |
| seed44 | step250, step270, step290, step310 | 226 (310: 224) |
vanilla_opd (vanilla on-policy distillation baseline)
| seed | contents | endpoint full274 |
|---|---|---|
| seed42 | step20…step240 every 20 steps + step250 (frozen-config b158 repro; per-step sweep) | 224 |
| seed43 | step250 | 216 |
| seed46 | step246 (explorer quota exhausted at 246) | 195 |
Not included: DeepSpeed optimizer states (needed only to resume training) and incomplete/duplicate exports. The A100 collaborator runs live on the collaborator's machine.
Base model license
Derivatives of Qwen/Qwen2.5-3B-Instruct (Qwen Research License) and of GiGPO-Qwen2.5-7B-Instruct-ALFWorld as teacher. Research use.