ryan-xman66/UOPD-3B-training-logs

Training logs at 3B: UOPD, OPD and TCOD-F2B (ALFWorld, WebShop)

Each run has two files: SLURM stdout (.out, the trainer's per-step metrics) and stderr (.err, launcher trace and library warnings).

file	benchmark / method	run	job, node, date	length
alfworld/uopd_3b_confidence_quantile_r030to005_beta1_takeover_job276698.{out,err}	ALFWorld, UOPD 3B (reported arm)	alfworld_sacd_3b_confidence_tau1.28_beta1.0_sftplain_lazy_teacher_takeover_saidConfQuantTakeover	276698, laniakea (4× A6000), 2026-08-17, 9 h 42 min	steps 1–250, all logged; final checkpoint saved; COMPLETED
webshop/opd_3b_job287103.{out,err}	WebShop, OPD 3B	webshop_opd_20260905021218	287103, voyager, 2026-09-05, 1 h 29 min	steps 1–150, all logged; final checkpoint saved; COMPLETED

UOPD settings, as the log's own rendered config records them: detector confidence, tau_mode: quantile, rate_schedule: linear with r_start 0.3, r_end 0.05, r_decay_steps 120, teacher_takeover: true, sft_coef (β) 1.0, sft_mode plain, student Qwen2.5-3B-Instruct, teacher GiGPO-Qwen2.5-7B-Instruct-ALFWorld.

ALFWorld 3B OPD and TCOD-F2B (June runs, alfworld/june_runs/). Both are complete 3B training runs of ours, trained on laniakea with the Qwen2.5-3B-Instruct student and the GiGPO-7B teacher. Each folder holds the run's own trinity logs (log/trainer.log, log/explorer.log, and the rest) and its TensorBoard event files (monitor/tensorboard/{trainer,explorer}/), which carry the per-step metrics.

folder	method	run	trainer steps	wall clock
alfworld/june_runs/opd_3b_full_lan/	OPD	opd_3b_full_lan	1–255, all logged	2026-06-02 20:00 → 06-04 00:56; resumed from its own checkpoint after a crash on 06-03, which is why there are three sets of event files
alfworld/june_runs/tcod_f2b_3b_full_lan_20260604005859/	TCOD-F2B	tcod_f2b_3b_full_lan_20260604005859	1–255, all logged	2026-06-04 01:04 → 09:55

Caveat. These runs predate a fix to the ALFWorld environment and evaluation harness. They are not the checkpoints behind the paper's ALFWorld OPD and TCOD-F2B rows, which evaluate released checkpoints (Wenboz/TCOD-v1-OPD-Qwen2.5-3B-ALFWorld and the released F2B step-250). Under the corrected evaluation, opd_3b_full_lan scores 0.43 / 0.30 on seen / unseen. Read them as training-dynamics logs, not as the paper's baselines.

Redaction. The UOPD .err echoed an HF access token on two lines. Both are replaced with hf_REDACTED. Nothing else was changed, and each file has the same line count as its original.