tt-x15-megatron — step 51 (X15 Megatron entropy rescue)

GRPO checkpoint from the TaskTrove X15 experiment, trained on Iris / CoreWeave (cw-rno2a, 4x8 H100, Megatron runtime). Base model Qwen/Qwen3-Coder-30B-A3B-Instruct, trained on DCAgent/exp_rpt_multifile with SkyRL + Terminus-2. X15 tested whether lr 8e-6 + eps_clip_low 0.3 + temp 1.2 prevents the X14 Megatron late collapse; the answer is negative — the arm collapsed the same way.

Checkpoint selection

global_step_51 — the owner-confirmed campaign selection (trailing-5 0.2199; step reward 0.2441; pass@8 0.4219; entropy 0.0194), the best pre-collapse saved checkpoint (near-tie with step 39 within smoothing flavor). Exported on-Iris from the sharded Megatron checkpoint bank (427.7 GB distcp, 30 objects) via the checkpoint-export job at the saved policy geometry (2x8 H100).

Run status — KILLED at 101/400 (COLLAPSED)

Reward peaked ~0.2273 trailing-5, collapsed to ~0 at steps 84/86 and held ~8% of peak for 13+ steps while tokens shrank 9.2k -> 7.6k. Mechanism (first identified for this family): entropy collapse shortened per-turn output until agents exhausted max_turns: 30 unfinished (cap-hitters 23% -> 100%, tokens/turn 572 -> 281). Not infrastructure. Not an admissible horizon result; the arm is preserved as the collapse-family negative control. Confound: three hyperparameters moved at once.

See training_logs/ for metrics.csv (stitched per-step curve + trailing-5 EMA), report.md, reward_plot.png, and the launch rl_config.yaml. W&B: dogml/OpenThoughts-Agent v5dzu6fw.

Training Traces

open-athena/tt-x15-megatron — 1/8 systematic subsample (every 8th trial of 53,411 = 6,677 selected, uniform coverage of steps 1-101). The full set (~256 GB) exceeded the sync host's free disk (47 Gi); 1/4 (~64 GB) also did not fit. Documented, owner-precedented deviation (same rule as tt-x10-fsdp2-fa2: smallest stride that fits).