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

records/track_non_record_16mb/2026-03-19_AdaptiveEvalContext_MLX_M4Pro_sp1024_200it/README.md

@@ -0,0 +1,57 @@
+This folder captures a non-record local MLX submission for adaptive eval-time context.
+
+The idea in this snapshot is simple: do one coarse pass over the validation stream, mark the harder windows from that pass, then rescore only those windows with a finer stride. The training setup stays close to the baseline MLX path; the change is in how the final roundtrip evaluation spends extra context.
+
+This is not a leaderboard claim. It is a local Apple Silicon result meant to document the idea, the code snapshot, and a same-setup comparison against standard final evaluation.
+
+Configuration:
+- Hardware: Apple M4 Pro, 48 GB unified memory
+- Track: non-record, local Apple Silicon MLX
+- Tokenizer/data: `fineweb10B_sp1024`, first train shard, first `32768` validation tokens
+- Model: SP-1024, `9x512`, `KV4`, tied embeddings
+- Training length: `200` iterations, `8192` train tokens/step
+- Final eval mode: adaptive
+- Adaptive eval settings: `coarse_stride=256`, `fine_stride=64`, `hard_fraction=0.25`
+
+Command used for the included adaptive run:
+```bash
+cd records/track_non_record_16mb/2026-03-19_AdaptiveEvalContext_MLX_M4Pro_sp1024_200it
+RUN_ID=cmp200_adapt_c256_f64_h025 \
+SEED=1337 \
+ITERATIONS=200 \
+TRAIN_BATCH_TOKENS=8192 \
+GRAD_ACCUM_STEPS=8 \
+VAL_LOSS_EVERY=0 \
+VAL_BATCH_SIZE=32768 \
+VAL_MAX_TOKENS=32768 \
+FINAL_ROUNDTRIP_EVAL=1 \
+FINAL_EVAL_MODE=adaptive \
+FINAL_EVAL_COARSE_STRIDE=256 \
+FINAL_EVAL_FINE_STRIDE=64 \
+FINAL_EVAL_HARD_FRACTION=0.25 \
+FINAL_EVAL_BATCH_SEQS=16 \
+DATA_PATH=../../../data/datasets/fineweb10B_sp1024 \
+TOKENIZER_PATH=../../../data/tokenizers/fineweb_1024_bpe.model \
+../../../.venv/bin/python train_gpt.py > train.log 2>&1
+```
+
+Included result (`train.log`):
+- Pre-quant eval at stop: `val_loss:4.1575`, `val_bpb:2.4070`
+- Post-roundtrip eval: `val_loss:4.15029331`, `val_bpb:2.40284524`
+- Eval time for final adaptive roundtrip pass: `2386ms`
+- Selected windows: `hard_windows:31/124`, `fine_windows:124`
+- Serialized model int8+zlib: `11239210 bytes`
+- Code size: `58701 bytes`
+- Total submission size int8+zlib: `11297911 bytes`
+
+Same-setup reference (`compare_standard.log`):
+- Standard final eval: `val_loss:4.16789573`, `val_bpb:2.41303630`
+- Eval time: `321ms`
+
+So in this local fixed-step proxy, the adaptive pass improves the final roundtrip score by about `0.01019 bpb` over the same setup with standard final evaluation, but it also increases final eval time. That tradeoff is the main reason this is being submitted as a non-record WIP rather than as a score claim.
+
+Included files:
+- `train_gpt.py` - exact MLX code snapshot used for the run
+- `train.log` - adaptive local run log
+- `compare_standard.log` - same-setup standard-eval comparison log
+- `submission.json` - metadata for the run