Add Deep14x416 KV2 non-record MLX submission (val_bpb=1.8440)

records/track_non_record_16mb/2026-03-19_Deep14x416_KV2_SP1024_MLX_750it/README.md

@@ -0,0 +1,55 @@
+This record captures an unlimited-compute non-record Apple Silicon MLX run using a deeper/narrower SP-1024 configuration.
+
+The main idea was to trade some width for depth while keeping the artifact comfortably under the `16,000,000` byte cap:
+- Layout: `VOCAB_SIZE=1024 NUM_LAYERS=14 MODEL_DIM=416 NUM_HEADS=8 NUM_KV_HEADS=2 MLP_MULT=2`
+- Trainer: root `train_gpt_mlx.py` snapshot copied into this record folder
+- Hardware: Apple Silicon (`Apple M5 Max`, local MLX run)
+- Dataset/tokenizer: published `fineweb10B_sp1024` export with `10` train shards downloaded locally plus the full fixed validation split
+- Track: non-record, unlimited compute, still under the `16,000,000` byte artifact cap
+
+This run also increases validation batch size and enables logit chunking so the full validation pass can complete in a reasonable amount of time on local hardware without changing the measured metric:
+- `VAL_BATCH_SIZE=8388608`
+- `LOGIT_CHUNK_TOKENS=65536`
+
+Command (run from this record folder so it uses the archived trainer snapshot):
+```bash
+cd records/track_non_record_16mb/2026-03-19_Deep14x416_KV2_SP1024_MLX_750it
+
+RUN_ID=deep14_416_kv2_full_750 \
+ITERATIONS=750 \
+MAX_WALLCLOCK_SECONDS=0 \
+TRAIN_BATCH_TOKENS=16384 \
+VAL_LOSS_EVERY=0 \
+VAL_BATCH_SIZE=8388608 \
+LOGIT_CHUNK_TOKENS=65536 \
+TRAIN_LOG_EVERY=50 \
+WARMUP_STEPS=10 \
+DATA_PATH=../../../data/datasets/fineweb10B_sp1024 \
+TOKENIZER_PATH=../../../data/tokenizers/fineweb_1024_bpe.model \
+NUM_LAYERS=14 \
+MODEL_DIM=416 \
+NUM_HEADS=8 \
+NUM_KV_HEADS=2 \
+MLP_MULT=2 \
+python3 ./train_gpt_mlx.py
+```
+
+Key metrics (from `train.log`):
+- Final pre-quant eval at step `750`: `val_loss:3.1118`, `val_bpb:1.8430`
+- Post-quant roundtrip eval: `val_loss:3.11359052`, `val_bpb:1.84404368`
+- Exact printed metric: `final_int8_zlib_roundtrip_exact val_bpb:1.84404368`
+- Train time: `529553ms` (`step_avg:706.07ms`)
+- Saved FP model: `63982252 bytes`
+- Serialized model int8+zlib: `12339367 bytes`
+- Code size: `49622 bytes`
+- Total submission size int8+zlib: `12388989 bytes`
+
+Training volume:
+- Global batch: `16384` tokens/step
+- Total train tokens seen: `12288000`
+- Local dataset subset: `10/195` train shards from `fineweb10B_sp1024`
+
+Included files:
+- `train_gpt_mlx.py` (exact code snapshot used for the run)
+- `train.log` (exact local training log)
+- `submission.json` (leaderboard metadata)

records/track_non_record_16mb/2026-03-19_Deep14x416_KV2_SP1024_MLX_750it/submission.json

@@ -0,0 +1,17 @@
+{
+  "author": "Clay Schubiner",
+  "github_id": "cschubiner",
+  "name": "Deep14x416 KV2 SP1024 MLX 750-step",
+  "blurb": "Unlimited compute track: a deeper/narrower Apple Silicon MLX run with 14 layers at width 416, 2 KV heads, SP-1024 tokenizer, and 750 training steps on a 10-shard FineWeb subset; final int8+zlib roundtrip scored 1.8440 BPB under the 16,000,000-byte cap.",
+  "date": "2026-03-19T06:34:39Z",
+  "track": "non-record-unlimited-compute-16mb",
+  "val_loss": 3.11359052,
+  "val_bpb": 1.84404368,
+  "pre_quant_val_loss": 3.1118,
+  "pre_quant_val_bpb": 1.8430,
+  "step_stop": 750,
+  "wallclock_seconds": 529.553,
+  "bytes_total": 12388989,
+  "bytes_model_int8_zlib": 12339367,
+  "bytes_code": 49622
+}