| { |
| "model": "D1A-E2B-it v0.6 (experimental)", |
| "base": "google/gemma-4-E2B-it", |
| "base_revision": "3e22461f65e89153144f8adb70e3b8c2cc9845a7", |
| "change_from_e2b_v0.6": "the trunk only: google/gemma-4-E2B (d29ff6b4) -> google/gemma-4-E2B-it (3e22461f); stage 2 uses E2B v0.6's mix and settings unchanged", |
| "stage_1": { |
| "recipe": "recipes/d1a-e2b.yaml (what made E2B v0.2.x) with base google/gemma-4-E2B-it", |
| "suite": "evals/v7/decision-v7", |
| "epochs": 2, |
| "lr": 0.0001, |
| "batch": 4, |
| "accum": 2, |
| "dtype": "bf16", |
| "weights_dtype": "bf16", |
| "p_none_pair": 0.25, |
| "checkpoint": "kept in the private repo JohnP1/d1a-e4b-runs (e2b-it/stage1), not published", |
| "weights_sha256": "ce9d3ac952af9c1045cea80e981bd0612b3efb9c199a308616886f77946bdcb7", |
| "head_sha256": "d4e3e8d0f07ee1b24c7fcca957052d91ea391b92ffb1aeddc88f838c70581806" |
| }, |
| "init_from": "stage 1 above (stage 2 starts from it, where E2B v0.6 started from E2B v0.2.1)", |
| "stage": "Every later skill in one run, as E2B v0.6 (jonpol01/d1a#229): PR labels (English 4,000, Japanese 753, blast radius 1,500, and 479 hermes-agent PRs created 2026-10-02 to 10-05 shown twice), hard-v1 4,000, devtools-v1 4,000, documents-v1 2,500, JGLUE 2,000, routing factory/generic 1,216/1,000, decision-v7 1,500; 23,427 records; 1 epoch, lr 5e-5, LoRA 16 on all projections, batch 2 x accum 4, states up to 5,120 tokens, 2,929 steps", |
| "release_note": "Published on the owner's decision of 2026-10-12 for an iOS proof of concept (an app that already loads Gemma 4 E2B-it), although the bar pre-registered for it (jonpol01/d1a#227, Bar B, 'good enough for apps') failed on two lines: sealed-EVAL severity -4.3 [-7.8, -1.0] against E2B v0.6 (needed >= -3) and media agreement with E4B v0.6 21/28 (needed >= 25/28). Experimental: E2B v0.6 stays E2B's model, and E4B v0.6 remains the PR labeler.", |
| "bos": "The Gemma 4 E2B-it tokenizer adds no <bos>, but the model was trained with one. d1a with jonpol01/d1a#261 prepends it (bos_token_id 2); clients that encode themselves must prepend id 2.", |
| "data": { |
| "counts": { |
| "pr-labels:fresh-fit": 958, |
| "pr-labels:train": 4000, |
| "pr-labels:train-ja": 753, |
| "pr-labels:train-blast": 1500, |
| "hard-v1:train": 4000, |
| "devtools-v1:train": 4000, |
| "documents-v1:train": 2500, |
| "ja-jglue:train": 2000, |
| "routing:factory-train": 1216, |
| "routing:generic-train": 1000, |
| "decision-v7:train": 1500 |
| }, |
| "records": 23427, |
| "sha256": "efff4b161e2d493f3d6e33712174614964279446ed2deb65b78ad4f7cb86e7a9", |
| "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 every eval partition" |
| }, |
| "training": { |
| "epochs": 1, |
| "lr": 5e-05, |
| "head_lr": 0.0, |
| "weight_decay": 0.01, |
| "lora": 16, |
| "lora_targets": "all", |
| "batch": 2, |
| "accum": 4, |
| "dtype": "bf16", |
| "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": "Hugging Face Jobs, one NVIDIA L4 (24 GB), stage 1 and stage 2 in one job", |
| "cost_usd_about": 3.51, |
| "finished": "2026-10-12 18:33 JST", |
| "stage_2_wall_seconds": 9914, |
| "stage_2_peak_device_bytes": 13770881024, |
| "optimizer_steps": 2929, |
| "records_seen": 23427, |
| "truncated_records": 0, |
| "rejected_records": 0 |
| }, |
| "calibration": { |
| "temperature": 1.414, |
| "fit": "pooled rows: decision-v7 calibration split + PR-labeling development set", |
| "questions": 3178, |
| "interval_90": [ |
| 1.35, |
| 1.481 |
| ], |
| "ece_raw": 0.053, |
| "ece_out_of_fold": 0.02, |
| "use_case_temperatures": "none" |
| }, |
| "code": "github.com/jonpol01/d1a (d1a.training.train, d1a.training.recipe): v0.6's commit b17f3077 plus the <bos> fix (branch e2b-it-bos, da9b5bfc; merged as jonpol01/d1a#261, b4b0f989)" |
| } |
|
|
