Record candidate: PR #1797 + AWQ-lite top3 + LQER 60k on b180-tlr56 — val_bpb 1.06043 (seed=0)

[vimeto]

Summary

Builds on the b180-tlr56 lineage (PR #1935) by stacking PR #1908's AWQ-lite mixed-precision GPTQ on top, plus a small LQER budget bump that uses cap margin freed by AWQ-lite's INT8 promotions.

Single seed (SEED=0): val_bpb 1.06043, val_loss 2.32062 nats. Beats PR #1855's 3-seed mean (1.06108 BPB, 2.32203 nats) by -0.00065 BPB / -0.00141 nats. Eval 599.3s, train 596.2s, both inside the 600s lane caps. Per-group lrzip artifact 15,947,372 bytes; total submission 15,982,182 (cap margin 17,818).

i'm running additional seeds at this configuration on RunPod 8xH100 right now and will append SEED=314 and SEED=1234 numbers as soon as they finish. some of those runs are exploring slightly different hparam neighborhoods (LQER budget, AWQ top_k) to map the local landscape — the headline single-seed value is from the configuration documented in the README.

This PR consolidates the test-plan items left open in PR #1935 (which promised SEED=0 / SEED=1234 multi-seed reference logs but couldn't complete them due to a pod crash on the original session). PR #1935 is being closed.

See records/track_10min_16mb/2026-05-05_AWQTop3_LQER60k_KLoRA_b180_tlr56/README.md for full recipe, ablation context, compliance notes, and lineage credits.

Test plan

• SEED=0 full retrain on H100
• Per-group lrzip artifact roundtrip verified lossless
• Total submission ≤ 16 MB
• SEED=314 at this configuration: running on RunPod 8xH100, will append
• SEED=1234 at this configuration: running on RunPod 8xH100, will append

[vimeto]

3-seed update is in (19576a74). SEED=314 → 1.06070; SEED=1234 → 1.06132.

Honest result: this PR is worse than the current merged SOTA.

Leaving this PR as draft for documentation of the AWQ-lite + LQER + drop-M LoRA stacking experiment and as the missing H100 multi-seed continuation of PR #1935. Not a record claim.