EXP-8 / completed / training
EXP-8: One-hour state-conditioned sanity run
Test whether state plus optional short history can support efficient next-move training in a fixed one-hour sanity budget.
Non-reasoning next-move model: explicit board state plus optional 16-ply history, with three supervised move tokens. No thought or search supervision. This differs from full-history CM-2 through CM-5. Follow data → tokenizer → training → harness → bot bridge.
Baseline
Warm start from CM-5: unchanged backbone and 128 old embedding rows, 384 new embedding rows; fresh AdamW. Not an optimizer resume.
Data
Sealed 10,000,000-position pool: 5,000,000 main-corpus and 5,000,000 curated-corpus positions. Consumed 3,426,304 distinct positions without replacement, exactly 1,713,152 from each source. This is a 50/50 position mix, not a game mix. Validation: 8,192 positions / 24,576 targets; no test evaluation.
Configuration
Exact recorded configuration
{
"run_id": "5e191684-ff0c-4b16-bf23-a6ab42c2daa7",
"parameters": 86235648,
"layers": 12,
"width": 768,
"heads": 12,
"vocabulary": 512,
"input": "FEN state plus optional 16-ply history",
"input_slots": 122,
"retained_position_rows": 1024,
"microbatch_per_gpu": 64,
"gpus": 2,
"accumulation": 4,
"decisions_per_update": 512,
"precision": "BF16 autocast; FP32 model/AdamW",
"optimizer": "fresh fused AdamW",
"selected_step": 6500
}Compute
Two RTX 4090 GPUs; 2026-10-29 17:42:39 to 18:40:12 UTC (57m32.830s). 6,692 updates, 10,278,912 prediction targets; 992.32 decisions/s all-in.
Results
| Metric | Value |
|---|---|
| Initial state-objective NLL | 10.82181 |
| Best state-objective NLL | 0.6436095 |
| Updates completed | 6,692 |
| Distinct positions | 3,426,304 |
| Prediction targets | 10,278,912 |
| All-in decisions per second | 992.3200 |
Validation learning curve (2 checkpoints)
| Update | Validation NLL |
|---|---|
| 0 | 10.821807 |
| 6500 | 0.643609 |
What this does not establish
- Different objective and validation units: NLL is not numerically comparable with full-history experiment losses.
- No playing-strength results and no promotion. A separate system-level evaluation is being prepared as EXP-9.
- The shorter input did not improve supervised decisions/s relative to the dense full-game objective. High GPU utilization alone is not useful-throughput improvement.
- Warm-start exposure remains a limitation. Last weights were not freshly validated; select the best checkpoint.
- Exact state-v2 source and reproduction packages are not yet in the public snapshot archive; shared backbone source is identified as lineage only.
Provenance
Shared backbone source lineage (not the state-v2 implementation) →
Checkpoint SHA-256 7b63bab559741893b632f456e5f9a8e48b15e48fc9c47c302734da9d613dfe85
Download experiment record (JSON)
Evidence: terminal-audit.json
Original artifact SHA-256 757114e47bb8db1f9896bb8db2b5e6ee9ad3b3c3c68b0af105efdf06cca4b96c
Selected public fields; private paths and process/account details excluded.
{
"run_id": "5e191684-ff0c-4b16-bf23-a6ab42c2daa7",
"no_promotion": true,
"best_checkpoint_sha256": "7b63bab559741893b632f456e5f9a8e48b15e48fc9c47c302734da9d613dfe85",
"last_checkpoint_sha256": "d5c9a22b67b809c7cb218637137e8fbce15a73217de75fe649faa835c42fc711",
"last_weights_freshly_validated": false,
"updates": 6692,
"decisions": 3426304,
"source_decisions_each": 1713152
}What this experiment validates
CM-6 is research-only and is not a promotion candidate. Its purpose is to validate an end-to-end pipeline for training and inference from FEN-derived current board/state plus optional recent history, without requiring the complete game history from the standard start.
FEN describes the current board and state; it is not an incomplete board position. The available past history can be incomplete, and repetition history can be unknown. Optional recent moves do not reconstruct the entire game.
The completed one-hour training run demonstrates that this input and target pipeline can learn. Legal game play tests the integrated model, authoritative legality handling and external mate helper; it does not establish the raw model’s legality or playing strength by itself. Limited training duration is context, not causal proof of why a particular score occurs. No duration ablation or controlled input-only comparison was performed.
EXP-9 uses an adaptive 1/3/5/7-ply helper, unlike the frozen three-ply helper in EXP-6 and EXP-7. Any comparison is system-level. CM-6 remained a research checkpoint; subsequent CM-7 adoption is documented separately.