Add classical 4-gram non-record submission

[Muhtasham]

Summary

This PR adds a non-record classical submission under records/track_non_record_16mb/2026-03-25_classical_4gram_10m_eval.

The submission is fully non-neural:

• discounted hashed 4-gram model with unigram/bigram backoff
• artifact built from the official fineweb10B_sp1024 train export
• exact evaluation on the official fineweb_val_* split
• no training-data access during evaluation; final eval uses the saved artifact via --load-state only

Exact Results

• val_bpb: 1.91070694
• validation tokens: 62,021,846
• evaluation wallclock: 571.97s
• model artifact bytes: 14,310,783
• code bytes: 57,801
• total bytes: 14,368,584

Included Files

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

Notes

This is a non-record submission, not a SOTA claim.

The final packaged line is the faster classical path:

• first 10,000,000 train tokens used to build the artifact
• --absolute-discount 0.75
• --ngram-contexts 3
• --mix-backoff-experts 0
• no cache experts
• no copy experts

I verified locally that:

• the artifact is under the decimal 16,000,000 byte cap
• the final full-validation run is performed on the official fineweb_val_* split
• the final evaluation does not read training shards
• the records folder compiles and runs from within the folder