WIP: Add adaptive eval-time context non-record MLX submission

[stpcoder]

Summary

This PR adds a non-record MLX submission under records/track_non_record_16mb/ for adaptive eval-time context.

The setup is a local Apple Silicon run, not a leaderboard claim. The main change in this snapshot is the final evaluation path: instead of using one fixed policy over the whole validation stream, it does a coarse pass first and then rescoring with a finer stride on the harder windows from that pass.

Included files

• README.md
• submission.json
• train.log
• compare_standard.log
• train_gpt.py

Run details

• hardware: Apple M4 Pro, 48 GB unified memory
• model: SP-1024, 9x512, KV4, tied embeddings
• train data: first FineWeb train shard
• validation: first 32768 validation tokens
• training length: 200 iterations
• adaptive eval settings: coarse_stride=256, fine_stride=64, hard_fraction=0.25

Result in train.log

• pre-quant eval at stop: val_bpb=2.4070
• final int8+zlib roundtrip: val_bpb=2.40284524
• total submission size: 11297911 bytes

Same-setup reference

compare_standard.log uses the same setup with standard final evaluation:

• standard final roundtrip: val_bpb=2.41303630
• adaptive final roundtrip: val_bpb=2.40284524

So in this local fixed-step proxy, the adaptive pass improves the final roundtrip score by about 0.01019 bpb, but it also makes the final eval pass slower. That tradeoff is the reason this is being submitted as a WIP non-record result rather than as a performance claim.

Validation

• python -m py_compile records/track_non_record_16mb/2026-03-19_AdaptiveEvalContext_MLX_M4Pro_sp1024_200it/train_gpt.py
• included local MLX run logs for both adaptive and standard final eval under the same setup

[MatoTeziTanka]

Community Review — WIP: Add adaptive eval-time context non-record MLX submission

Compliance: NEEDS AUTHOR ACTION — train_gpt.py fails to import on CT2038 (Python 3.10 / torch 2.10.0+cpu)

What I found: The CPU smoke test on CT2038 (proteus-engine, 128 GB RAM, Triton 3.6.0, flash_attn stub, cutlass_evt_fusion stub) failed at the import step with:

ModuleNotFoundError: No module named 'mlx'

A few of the common patterns I've seen for this class of error in the 2026-04-11 sweep:

• PEP 701 f-string nesting — e.g. log(f" {cat}: {", ".join(...)}") is valid Python 3.12+ but invalid Python 3.10 because the inner ", " re-enters the outer double-quote context. One-character fix: ', ' instead of ", ". See PR Record: SP8192 + Improved Parallel Residuals + Muon 0.97 + LR 0.03 + Legal TTT — val_bpb 1.07785 (3-seed mean) #1541 / Record: SP8192 + Triple Recurrence + Banking + Fused MLP + Muon 0.97 — val_bpb 1.0778 (3-seed mean) #1523 for reference.
• Missing flash_attn variants — e.g. from flash_attn_interface import flash_attn_varlen_func when the wrapper script only stubs flash_attn_func. Not a PR defect on H100s, but the eval image / CPU preflight path needs a guarded import.
• Local compiled extension — e.g. import cutlass_evt_fusion from a records/*/cutlass_evt_fusion/ subfolder that isn't on the import path at smoke time. Usually an import-order issue inside the script.
• Actual syntax error — typo, missing bracket, etc.

Recommendation: Could you run python3 -c "import py_compile; py_compile.compile('train_gpt.py')" on your records-folder train_gpt.py under Python 3.10 specifically? The eval image is Python 3.10 per Issue #17 / the README, so any parse error on 3.10 blocks the submission at import time before any of the scored-eval logic runs.

Once the parse/import issue is fixed, I'll re-run the compliance audit through the normal pipeline. No other flags identified yet because the audit halts at the import step.

---

Reviewed by @MatoTeziTanka — The Agora. CPU smoke test (CT2038 proteus-engine, 2026-04-11): IMPORT_FAIL — ModuleNotFoundError: No module named 'mlx'. Classification via classify_prs.py AST-based classifier; full compliance audit deferred until the import issue is resolved. Auto-drafted from a template and spot-checked before posting.