Summary

mean val_bpb=1.1642 across 3 seeds on 8xH100 | Artifact: ~15.85 MB (under 16MB)

Six improvements stacked on the baseline:

1. Vocab 4096 — custom SentencePiece BPE tokenizer (included). More bytes per token = better BPB.
2. 3x MLP expansion (hidden=1536) — enabled by int6 quantization savings.
3. Int6 STE fake quantization — quant gap only +0.005 BPB.
4. Seq4096 training — 4x longer context.
5. SWA — 7 checkpoint average during warmdown.
6. Sliding window eval (stride=256) — 148s eval time on 8xH100.

3-Seed Validation

Mean: 1.1642, Std: 0.0007, p < 0.0001 (one-sample t-test vs baseline 1.2244)

Command

VOCAB_SIZE=4096 NUM_LAYERS=8 TRAIN_SEQ_LEN=4096 MLP_MULT=3 \
WARMDOWN_ITERS=3000 WEIGHT_QUANTIZATION_BITS=6 EMBED_QUANTIZATION_BITS=8 \
EVAL_STRIDE=256 SWA_ENABLED=1 MATRIX_LR=0.02 SCALAR_LR=0.02 TIED_EMBED_LR=0.03 \
MUON_MOMENTUM=0.99 MUON_MOMENTUM_WARMUP_START=0.92 MUON_MOMENTUM_WARMUP_STEPS=1500 \
MAX_WALLCLOCK_SECONDS=600 \
DATA_PATH=./data/datasets/fineweb10B_sp4096 \
TOKENIZER_PATH=./data/tokenizers/fineweb_4096_bpe.model \
torchrun --standalone --nproc_per_node=8 train_gpt.py

Tokenizer and dataset at sproos/parameter-golf-tokenizers.

Files

• train_gpt.py — self-contained script (1390 lines)
• train.log + train_seed1337.log, train_seed42.log, train_seed7.log
• submission.json, fineweb_4096_bpe.model