Summary

• Supersedes Record: 0.0498 bpb - Packed Training N-gram Artifact + Learned Weighting Gate (updated)  #931 for this line of work with the final low eval-time memory regime: the packed order-2..9 training n-gram artifact and learned gate remain, but the logistic context mixer and long phrase cache are removed from the final eval path.
• Final 3-seed mean val_bpb is 0.02139943 +/- 0.00003918; worst-case total submission size is 15,881,331 bytes and worst-case eval time is 437s.
• All reported runs stay within budget: training <600s, eval <600s, artifact <16MB.
• The main auxiliary eval-time state is the fixed 2 MiB order-2..9 n-gram cache: 32K buckets with two uint32 count tables per order. This is the primary persisted state beyond the transformer itself and it does not grow with validation length.
• The submission keeps the compliant causal path: the n-gram cache persisted from training time is included as part of the artifact itself, expert availability is context-only, GPTQ calibration uses cached training batches, the output distribution is normalized to sum to 1 for each token, and the reported path uses TTT_EPOCHS=0.

Results

3-seed mean val_bpb: 0.02139943

Sample std: 0.00003918

Causal Inference Scheme

1. Deserialize the packed order-2..9 n-gram cache from the submitted artifact at eval start.
2. Score each validation chunk once using only left context and the current cache state.
3. Query n-gram experts using left context only; the learned gate's expert-availability mask depends only on context evidence.
4. Blend neural + n-gram experts, then renormalize the full-vocabulary distribution so it sums to 1 before scoring.
5. Update the streaming n-gram cache only after the chunk has already been scored.
6. Report the final single-pass path with TTT_EPOCHS=0.

Compliance

• This is not a 2-pass method.
• Validation is scored in a single causal pass: each chunk is scored before that chunk is used for any cache update.
• The warm-start cache used at eval step 0 is part of the artifact itself, not a separate runtime input.
• The n-gram cache persisted from training time is included as part of the artifact and deserialized at eval start.
• The packed n-gram cache in the artifact is derived from training data only and is produced within the 600 second training budget.
• The learned gate does not use the true next token to decide which experts are available.
• GPTQ calibration runs inside the reserved pre-export budget using cached training batches from the same timed run; it does not reopen training shards after the official wallclock limit.
• The output distribution is normalized to sum to 1 for each token before likelihood is accumulated.
• The current reported numbers use TTT_EPOCHS=0, so there is no backward test-time adaptation in the final submission path.
• No future validation tokens are visible when scoring the current chunk.

Reproduction

pip install -r records/track_10min_16mb/2026-03-27_LowEvalMemoryRegime_PackedTrainCache_NoMixer/requirements.txt

cd records/track_10min_16mb/2026-03-27_LowEvalMemoryRegime_PackedTrainCache_NoMixer

SEED=1337 \
DATA_PATH=/root/parameter-golf/data/datasets/fineweb10B_sp1024 \
TOKENIZER_PATH=/root/parameter-golf/data/tokenizers/fineweb_1024_bpe.model \
ARTIFACT_NGRAM_EXPORT=1 \
MAX_WALLCLOCK_SECONDS=600 \
VAL_LOSS_EVERY=0 \
USE_MIXER=0 USE_PHRASE_CACHE=0 MIXER_HEAD=multi \
USE_NGRAM_CACHE=1 NGRAM_EVAL_ORDER=9 \
TRAIN_ORACLE_BUCKETS=32768 NGRAM_EVAL_BUCKETS=32768 \
USE_REGIME_TRACKER=0 USE_LOGIT_CAL=1 \
TTT_EPOCHS=0 TTT_FREEZE_BLOCKS=2 TTT_LR=0.0001 \
TTT_CHUNK_TOKENS=131072 SKIP_SLIDING=1 EVAL_STRIDE=64 TTT_TEMPERATURE=0.85 \
CROWN_Q_LAMBDA=0.01 PRUNE_PCT=0.05 BIGRAM_VOCAB_SIZE=0 \
GPTQ_CALIBRATION_SEQS=128 \
RENORMALIZE_FINAL_PROBS=1 VERIFY_FINAL_PROBS=1 \
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
torchrun --standalone --nproc_per_node=8 train_gpt.py

Submission Checklist

• One new folder added under records/track_10min_16mb
• README.md included
• submission.json included
• train_gpt.py included
• Train logs included (train_seed1337.log, train_seed42.log, train_seed7.log)
• Train and eval under 10 minutes
• Artifact under 16MB
• No tokenizer/dataset edits
• Score-first ordering preserved (no hindsight path)