Non-record: 11L GEPA + 12k Steps + Pure Int6 + Legal TTT (val_bpb=1.1079)

[Christopher-Lee-McClendon]

Non-Record Submission: 11L GEPA + 12k Steps + Pure Int6 + Legal TTT

val_bpb = 1.1079 (1.10788263 exact) | Pre-TTT float: 1.1268 | Int6 quant: ~1.154 | TTT gain: −0.046 | Artifact: 14.79 MB

Summary

GEPA architecture (11L, 27M params) trained for 12,000 steps (7k peak-LR + 5k warmdown) on 4×A100-40GB. Pure int6 per-row quantization with 15-candidate GPTQ-lite clip search + zstd-22 compression. Legal score-first TTT (SGD, momentum 0.9, lr=0.002, 10 epochs, freeze first 2 blocks).

Key Results

Metric	Value
Final val_bpb	1.10788263
Pre-TTT float (step 12000)	1.1268
Post-quantization pre-TTT	~1.154
TTT improvement	−0.046
Model bytes	15,432,359
Code bytes	78,281
Total artifact	15,510,640 (< 16 MB ✅)
Training time	~100 min (4×A100)
Eval time	2072s

Novel Contributions

1. 12k-step training with 5k-step warmdown — exploits unlimited-compute track
2. Pure int6 per-row quantization (no mixed int6/int8) with 15-percentile GPTQ-lite
3. Legal score-first TTT with LR warmup and SGD momentum

Track

track_non_record_16mb — unlimited compute, 16 MB artifact limit.

Checklist

• README.md with approach description
• submission.json with correct metadata
• train.log demonstrating results
• train_gpt.py (self-contained, no network calls)
• Total submission ≤ 16 MB
• Legal TTT (score-first, no val data cheating)
• GPG-signed commit

[MatoTeziTanka]

Community Review — Non-record: 11L GEPA + 12k Steps + Pure Int6 + Legal TTT (val_bpb=1.1079)

BPB: 1.1079 | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1416/#1423 pattern)

What I found in the code (head SHA 3bf3fefe03bf, file records/track_non_record_16mb/2026-03-24_11L_GEPA_12kSteps_PureInt6_LegalTTT/train_gpt.py):

The TTT path at line 399 implements the score-first-per-chunk pattern: each chunk is scored under torch.no_grad() / inference_mode() before the base_model.train() + SGD adaptation runs on that same chunk, with an is_last_chunk guard so the final chunk gets no adaptation pass. This is the structural shape the legal frontier uses (PRs #1416 erichroepke, #1423 aryanbhosale).

Per Issue #402 and Issue #677, TTT is legal when each token is scored before the adapter updates on it, and that's what the code does here — chunk ci is scored under weights adapted only on chunks 0..ci-1. No prequant_ttt_adapt_adamw(val_tokens, ...) multi-epoch fine-tune, no scored-region SLOT, no target-in-key n-gram cache.

CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 0.03s, dim=512, layers=11, vocab=1024, code=78281 B, SMOKE_TEST_PASS

Verdict: LOOKS CLEAN.

Recommendation to @cocohearts @valerio-oai @0hq @yuzhougu-oai @notapplica: MERGE pending standard checks (3-seed validation, 16MB artifact cap, 10-min wallclock on 8×H100 SXM). The compliance picture matches the legal reference frontier and no flags were raised by the classification pass.

Auto-classification caveat: this review was drafted by the AST-based classifier against a template derived from manually-reviewed cluster PRs (#1420, #1450, #1487, #1541, #1529, #1533, #1518). If I've misread a subtlety in your eval path — e.g., multi-epoch TTT that I mistook for single-pass, or a target-in-key lookup I missed in a helper function — please flag it and I'll re-run the audit manually.

---

Reviewed by @MatoTeziTanka — The Agora. CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 0.03s, dim=512, layers=11, vocab=1024, code=78281 B, SMOKE_TEST_PASS. Classification via deterministic AST-based classify_prs.py (pattern bank derived from ~65 manually-reviewed PRs earlier in the 2026-04-11 sweep). This review was auto-drafted from a template and spot-checked before posting — if the template misread your code, please call it out so I can iterate the classifier.