Full GPTQ + XSA-all + SWA/EMA + Score-First TTT (compliant BPB=1.1175, 1-seed)

[Robby955]

Results (Compliant)

Config	Seed	Sliding BPB	Artifact	Train Time	Status
Score-first TTT (3 epochs)	1337	1.1175	15.87 MB	560s train + 40s GPTQ = 600s	Compliant
No TTT	1337	1.1182	15.93 MB	560s train + 40s GPTQ = 600s	Compliant

Best compliant score: 1.1175 BPB (single seed 1337, 3-seed validation in progress)

Note: Previous headline numbers (1.1158, 3-seed mean 1.1163) were from runs where GPTQ calibration ran outside the 10-minute training budget. Those numbers have been retracted per the organizer ruling in Issue #677. All figures above use compliant timing where GPTQ calibration is counted within the training budget.

---

Compliance

Training time budget (600s total on 8×H100 SXM):

• 560s: Main training loop (Muon + Adam, ~6,230 steps)
• 40s: GPTQ int6 quantization (calibration on training data)
• Total: 600s (within 10-minute budget)

Test-time training (TTT):
Score-first protocol — the model scores each validation chunk before adapting on it. No token is ever re-scored after adaptation. This follows the causal/streaming TTT pattern confirmed legal by organizers (Issue #402, #677). Full-epoch TTT (train on all val data before scoring) is NOT used.

Artifact: 15.87 MB (code: ~94KB, compressed weights: ~15.78MB). Under 16,000,000 byte limit.

---

Key Contributions

1. Full GPTQ halves the quantization gap (0.008 → 0.004 BPB)

Cholesky-based GPTQ with act-order column permutation and block-wise error compensation (block_size=128). Calibrated on 256 training batches with 1% diagonal damping.

Method	Quant Gap	Artifact Size
Naive int6 (grid search)	+0.0083 BPB	15.81 MB
Full GPTQ (Cholesky)	+0.0039 BPB	15.92 MB

2. XSA on all 11 layers

Cross-Sequence Attention on all transformer layers (vs last 4 in baseline). Provides extended context beyond training sequence length at eval time. Worth ~-0.0013 BPB.

3. SWA/EMA weight blending

Stochastic Weight Averaging over the final warmdown phase (16 snapshots every 50 steps), blended 50/50 with EMA (decay=0.997). Smooths weight landscape before quantization.

4. Score-first TTT (legal)

Sequential online adaptation: score chunk, then train on it with AdamW (lr=1e-4, 3 epochs, freeze first 9/11 blocks, 128K token chunks). Improves sliding BPB by -0.0007.

Architecture & Training

• Base: PR Record: Late Soft-Round QAT + Score-First Backward-Looking TTT — val_bpb 1.1178 #589 (11L GEPA, 8 GQA heads / 4 KV, VE128, Late QAT)
• Changed: XSA all 11 layers, Full GPTQ, SWA+EMA 50/50 blend, warmdown 4000
• Training: ~6,230 steps at 89.9ms/step on 8×H100 SXM, 27.1M params
• Artifact: 15.87 MB (code + lzma-compressed int6 weights)

Credits

• @RoyiRa for PR Record: Late Soft-Round QAT + Score-First Backward-Looking TTT — val_bpb 1.1178 #589 base architecture
• @thwu1, @unnir, @JoeProAI for components adopted in Record: Late Soft-Round QAT + Score-First Backward-Looking TTT — val_bpb 1.1178 #589
• @saml212, @kshitizz36 for GPTQ reference implementations (Non-record: 11L XSA-all + Full GPTQ + Selective Pruning (val_bpb=1.1154, 3-seed) #609, Record: Full GPTQ + LeakyReLU² + Parallel Muon (3-seed mean 1.1180) #626)
• @CiprianFlorin-Ifrim for int5+Soft-Round QAT reference (Record: int5 GPTQ + Soft-Round QAT (3-seed mean 1.1162) #606)
• @raahilshah for PR Record: 11L XSA-all + Full GPTQ (Budget-Legal) + Parallel Muon + Selective Pruning (val_bpb: 1.1178, 3-seed mean) #634 (XSA-all, selective pruning insights)

🤖 Generated with Claude Code

[Robby955]

Compliance Update

After reviewing the organizer ruling on PR #606 (GPTQ calibration must fit within the training time budget), I've re-run with compliant timing:

Training ends at 560s, GPTQ calibration runs at 560-600s, within the 600s training budget.

Compliant Results (560s training + 40s GPTQ calibration = 600s total)

Config	Seed	Pre-quant	Sliding BPB	Artifact
Compliant + TTT	1337	1.1374	1.1175	15.87 MB
Compliant, no TTT	1337	1.1375	1.1182	15.93 MB

TTT config (matching PR #606's recipe): AdamW lr=1e-4, 3 epochs, zero weight decay, freeze first 9/11 blocks, 128K token chunks.

The original 1.1158 score was from a 600s training run where GPTQ calibration ran outside the training budget — the same issue that affected PR #606. The compliant score with the same architecture is 1.1175.

3-seed validation on the compliant config is in progress. Will update the PR title and body once complete.

[valerio-oai]

This submission doesn't have a compliant /records submission (no submission.json or train logs), and it is using GPTQ-training-data calibration at eval time, which is disallowed. Closing for now.