Record: 11L EMA + Int6 + XSA + LeakyReLU² + Partial RoPE (val_bpb: 1.1309)

[parinzee]

Summary

• val_bpb: 1.1309 (mean of 3 seeds, std: 0.00017)
• 11 layers, 512 dim, 8H/4KV GQA
• Artifact: ~15.8 MB (all seeds under 16 MB)

3-Seed Results

Seed	val_bpb	artifact_bytes
42	1.13109	15,764,564
1337	1.13085	15,626,741
2024	1.13067	15,923,256
Mean	1.13087
Std	0.00017

Key Changes from Baseline

1. 11 layers (up from 10), 512 dim, 8 heads / 4 KV heads (GQA)
2. XSA (Exclusive Self Attention) on last 4 layers for better representation
3. LeakyReLU(0.5)² activation — squared leaky ReLU with 0.5 negative slope
4. Partial RoPE — only 16/64 dims use rotary embeddings
5. EMA weight averaging (decay=0.997) for smoother final weights
6. Int6 quantization for all large weight matrices + zstd-22 compression
7. Scale clamping fix — clamp_min(1/clip_range) improves quantization quality
8. Smaller batch size (524288 tokens) to fit more training steps (~8200 steps in 600s)
9. BigramHash(2048, dim=128) token embeddings
10. warmdown_iters=4500 for learning rate schedule
11. Higher learning rates (matrix_lr=0.025, scalar_lr=0.025)

Run Command

torchrun --standalone --nproc_per_node=8 train_gpt.py

Built on SOTA baseline by @thwu1 (PR #180).

[MatoTeziTanka]

Community Review — Record: 11L EMA + Int6 + XSA + LeakyReLU² + Partial RoPE (val_bpb: 1.1309)

BPB: 1.1309 | Compliance: LOOKS CLEAN — pure-neural submission, no TTT/SLOT/n-gram-cache

What I found in the code (head SHA 739a729f6b4a, file records/track_10min_16mb/2026-03-23_11L_EMA_Int6_XSA_LeakyReLU_PartialRoPE/train_gpt.py):

Static code review found no TTT adaptation function, no SLOT optimization loop, no n-gram-cache class, and no pre-quant val-token fine-tune. The eval path uses the standard sliding-window stride-64 pattern. The submission is a pure-neural architecture iteration on the standard SP1024/SP4096/SP8192 baseline.

CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 0.03s, dim=512, layers=10, vocab=1024, code=53603 B, SMOKE_TEST_PASS

Verdict: LOOKS CLEAN.

Recommendation to @cocohearts @valerio-oai @0hq @yuzhougu-oai @notapplica: MERGE pending the usual record-track checks (3-seed validation, under-16MB artifact cap, ≤600s train + ≤600s eval on 8×H100 SXM). No compliance flags from the classification pass — this looks like a clean pure-neural iteration on the standard baseline.

Auto-classification caveat: this review was drafted by the AST-based classifier. If there's a non-standard eval mechanism (logit postprocessing, hedge mixing, etc.) that I missed because it's factored into a helper file or a non-standard function name, 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=10, vocab=1024, code=53603 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.