| { |
| "model": "D1A-E4B v0.6", |
| "base": "google/gemma-4-E4B", |
| "base_revision": "411aa17b749aa952df1359d2dcea73917a544d9a", |
| "init_from": "JohnP1/d1a-e4b@v0.5 (commit e7db32f2254b6feb6ec830ae98662755dce39504)", |
| "stage": "Current PR labels (D3-fresh, jonpol01/d1a#198): 479 hermes-agent PRs created 2026-10-02 to 10-05, shown twice (one pass in each half of training); documents-v1 2,520 (never trained before); routing factory/generic 600/600; decision-v7 1,000; hard-v1, devtools-v1, JGLUE, PR labels 300 each; PR labels Japanese/blast 100/100; 7,078 records; 1 epoch, lr 2e-5, LoRA 16 on all projections, batch 2 x accum 4, states up to 5,120 tokens, 885 steps", |
| "release_exception": "Ships despite decide.py's veto V1 against v0.5 (hard-v1 -2.7 [-4.9, -0.5]), by the owner's decision of 2026-10-10: the model's main job is labelling current PRs, where v0.6 is better. Pin @v0.5 for hard-v1-style reasoning or the older PR-label conventions.", |
| "data": { |
| "counts": { |
| "pr-labels:fresh-fit": 958, |
| "pr-labels:train": 300, |
| "pr-labels:train-ja": 100, |
| "pr-labels:train-blast": 100, |
| "hard-v1:train": 300, |
| "devtools-v1:train": 300, |
| "documents-v1:train": 2520, |
| "ja-jglue:train": 300, |
| "routing:factory-train": 600, |
| "routing:generic-train": 600, |
| "decision-v7:train": 1000 |
| }, |
| "records": 7078, |
| "sha256": "31f7167055468b078e4c8ec16a4ce507fbc6e4662b85effa8bbe8a01dd8477e1", |
| "seed": 0, |
| "fresh_prs": { |
| "prs": 479, |
| "passes": 2, |
| "created": "2026-10-02 to 2026-10-05", |
| "text": "private; aggregates only" |
| }, |
| "leak_checks": "0 eval-state overlap, 0 PR-id overlap with the sealed fresh EVAL half and 47 eval partitions" |
| }, |
| "training": { |
| "epochs": 1, |
| "lr": 2e-05, |
| "head_lr": 0.0, |
| "weight_decay": 0.01, |
| "lora": 16, |
| "lora_targets": "all", |
| "batch": 2, |
| "accum": 4, |
| "dtype": "fp32", |
| "weights_dtype": "bf16", |
| "checkpointing": 1, |
| "max_state": 5120, |
| "p_none": 0.1, |
| "p_none_distract": 0.12, |
| "p_distract": 0.15, |
| "seed": 0 |
| }, |
| "compute": { |
| "where": "local, one Apple M1 Max (MPS)", |
| "wall_seconds": 60962, |
| "optimizer_steps": 885, |
| "records_seen": 7078, |
| "truncated_records": 0, |
| "rejected_records": 0 |
| }, |
| "calibration": { |
| "temperature": 1.954, |
| "fit": "pooled held-out rows: decision-v7 calibration split + PR-labeling development set", |
| "ece_raw": 0.093, |
| "ece_out_of_fold": 0.015, |
| "use_case_temperatures": "none in v0.6; a refitted routing temperature is planned for v0.6.1" |
| }, |
| "code": "github.com/jonpol01/d1a (d1a.training.train, scripts/quality_gate.py, scripts/decide.py)" |
| } |
|
|
