Record: 11L Int5-All + XSA5 + EMA + 10% Pruning (val_bpb=1.1466)

[trasnake87]

11L Int5-All + XSA5 + EMA + 10% Pruning

val_bpb = 1.1466 (sliding window stride=64) | 14.8 MB artifact | 8xH100 SXM, 600s

What's new

Two novel modifications on top of the XSA+EMA stack (PR #315):

1. Uniform int5 quantization for both MLP and attention weights. Prior SOTA uses int5 for MLP and int6 for attention. Through systematic post-training quantization search, we found attention weights tolerate int5 as well as MLP weights. Saves ~1MB compressed.

2. 10% magnitude pruning after EMA averaging, before quantization. Zeroes the smallest weights, which compress to nearly nothing under zstd. Saves ~500KB.

Combined: artifact drops from ~15.6MB to 14.8MB with minimal quality impact.

Results (seed 42, 8xH100 SXM)

Metric	Value
Steps	5,987
Pre-quant val_bpb	1.1504
Post-quant standard	1.1703
Post-quant sliding s64	1.1466
Artifact	14,811,335 bytes

Run

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

All defaults baked into the script. No env vars needed.

Built on PR #315 by @alertcat.

[ElliotSlusky]

not a record

[MatoTeziTanka]

Community Review — Record: 11L Int5-All + XSA5 + EMA + 10% Pruning (val_bpb=1.1466)

BPB: 1.1466 | Compliance: LOOKS CLEAN — pure-neural submission, no TTT/SLOT/n-gram-cache

What I found in the code (head SHA 2a18890025a7, file records/track_10min_16mb/2026-03-22_11L_Int5All_XSA5_EMA_Prune10/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 0.05s, dim=512, layers=11, vocab=1024, code=69722 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 0.05s, dim=512, layers=11, vocab=1024, code=69722 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.

[trasnake87]

Thanks for the thorough review, @MatoTeziTanka! Confirming the submission is pure-neural — no TTT, no SLOT, no n-gram cache, and the eval path is the standard sliding-window stride-64 pattern.

On the 3-seed validation requirement: we ran seed=42 and got val_bpb=1.1466 on 8×H100 SXM within the 10min + 10min budget. We'd be happy to run seeds 1337 and 2024 for statistical significance if maintainers want to consider this as a record candidate — just let us know.