Record: PR #1908 reproduction with compliant 600s wallclock — val_bpb 1.06044 (3-seed mean)

[AayushBaniya2006]

Summary

3-seed mean val_bpb = 1.06043952 (std 0.00091) on top of @romeerp's PR #1908 stack, run with organic 600 s wallclock control instead of FORCE_STOP_STEP=4945. All three seeds finish strictly under the 600,000 ms training cap.

Seed	Train wallclock	Post-TTT val_bpb	Artifact bytes
42	599,521 ms	1.05938494	15,943,518
0	599,665 ms	1.06101359	15,945,548
1234	599,676 ms	1.06092004	15,950,342
Mean	599,621 ms	1.06043952	15,946,469

vs PR #1908

	PR #1908	This submission	Δ
3-seed mean post-TTT	1.06081076	1.06043952	−0.00037124
Seed 42 train wallclock	601,153 ms (over cap)	599,521 ms	−1,632 ms
Max artifact bytes	15,996,559	15,950,342	−46,217

PR #1908 used FORCE_STOP_STEP=4945, which causes train_gpt.py to ignore the wallclock cap and forces step 4945 regardless of elapsed time. Their seed-42 run consumed 601,153 ms — over the 600,000 ms cap that defines the track_10min_16mb track. This submission removes FORCE_STOP_STEP and lets training stop organically at the wallclock cap; on the GPU instance used here (8×H100 80GB SXM, RunPod community cloud) that yields stop steps within 15 of PR #1908's targets, while staying compliant.

The PR #1908 author flagged this themselves: "this should be open for anyone who wants to claim a new record if they can re-run it under the 600s wallclock" (source). This submission takes them up on it.

What changed

train_gpt.py is byte-identical to PR #1908's submission file (commit 291d3abd on romeerp/parameter-golf:codex/awq-stepmatched). All architectural and quantization logic — including activation-aware GPTQ mixed-precision, LQER asymmetric int4, and per-group lrzip-zpaq compression — is unchanged. All quantization knobs (AWQ_LITE_GROUP_TOP_K=1, LQER_TOP_K=3, LQER_GAIN_SELECT=0) are unchanged.

The only difference is the wallclock control path:

	PR #1908	This submission
MAX_WALLCLOCK_SECONDS	0	600
FORCE_STOP_STEP	4945	unset

Reproducing

Same dataset and tokenizer as PR #1908: romeerp/parameter-golf-caseops-v1 (HuggingFace), variant sp8192_lossless_caps_caseops_v1_reserved. See records/track_10min_16mb/2026-04-29_PR1908Repro_Compliant600s_VAL_BPB_1.06044/README.md for the full env-var block.

Credits

This stands entirely on the work of @romeerp (PR #1908, PR #1729), @codemath3000 (PR #1855), @dexhunter (PR #1797, PR #1626), @nprime06 (PR #1787), and the rest of the PR #1855 lineage. The contribution of this submission is narrow: demonstrating that the PR #1908 stack achieves its quality strictly within the 600 s training cap when run with organic wallclock control.

Test plan

• All three seeds run with MAX_WALLCLOCK_SECONDS=600 and no FORCE_STOP_STEP
• All three train_seed*.log files captured and included
• All three artifacts under 16,000,000 bytes
• All three training wallclocks under 600,000 ms
• submission.json includes per-seed metadata + compliance attestation

🤖 Generated with Claude Code