Add non-record 1xH200 fp16-embed baseline sweep submission

[itu-itis24-buyukhelvacigilm24]

Summary

This PR adds a non-record submission that captures the best completed result from a controlled 1xH200, 10-minute screening sweep on the Parameter Golf baseline family.

The submitted run keeps the published SP-1024 baseline layout and only changes export behavior by preserving tok_emb.weight in fp16 during int8 export. In our H200 sweep, this was the strongest completed branch under the 16,000,000 byte cap.

What is included

• exact train_gpt.py snapshot used for the run
• exact train.log
• submission.json
• README.md with configuration, metrics, and sweep context

Why non-record

• this run was screened on 1xH200, not yet verified on 8xH100
• it does not claim a new leaderboard record
• it is intended as a reproducible, baseline-adjacent result and an honest report of ablations that did and did not help

Key result

• final_int8_zlib_roundtrip_exact val_bpb: 1.32078403
• total artifact size: 14,327,135 bytes

Sweep context

Variant	Exact final val_bpb
Clean baseline	1.32171904
FP16 embed passthrough	1.32078403
Sink4	1.32107101
MTP1 fixed	1.33126842
MTP2 fixed	1.32792538
FP16 embed + MTP2 fixed	1.32718519

Takeaway

In this budget, a small exporter-side precision change beat the more ambitious objective-side changes we tested. Our next step is to validate this baseline-adjacent fp16 embedding path on official 8xH100 hardware rather than continuing to add training complexity.