D1A-E2B-it v0.6 · MLX 8-bit (experimental)

A small decision model for Apple Silicon: give it a document and a few questions, and it picks the answers. Each question is yes/no, multiple choice or a score; D1A returns a calibrated probability for every option in one forward pass, with no generated text, so every answer is a valid option.

This is an experimental build for on-device apps that already load Gemma 4 E2B-it. It is D1A-E2B v0.6's recipe on the google/gemma-4-E2B-it trunk. Where it has been measured it is below JohnP1/d1a-e2b-mlx-q8 v0.6 (severity −4.3 points, media 21/28 against 25/28 agreement with E4B v0.6; numbers below), and E4B v0.6 (JohnP1/d1a-e4b-mlx-q8) remains the PR labeler. If your app does not already load Gemma 4 E2B-it, use E2B v0.6.

from d1a import D1A   # pip install "d1a[serve,mlx,media] @ git+https://github.com/jonpol01/d1a"  (needs jonpol01/d1a#261)
m = D1A.load("JohnP1/d1a-e2b-it-mlx-q8@v0.6")
m.decide("The customer was charged twice for the same order and wants a refund today.",
         {"team": {"type": "choice", "instructions": "Which team handles this?",
                   "criteria": {"billing": None, "shipping": None, "support": None}},
          "angry": {"type": "noul", "instructions": "Is the customer angry?"}})
# -> {"team": {"choice": ..., "probabilities": {"billing": ..., "shipping": ..., "support": ...}}, "angry": {"noul": ...}}

Not a chat model. mlx_lm.generate and LM Studio will load the Gemma 4 backbone and write ordinary text, but D1A's answers come from its decision head, which only the d1a package uses (see Run it).

<bos>: encode with it. The Gemma 4 E2B-it tokenizer adds no <bos>, but this model was trained with one at the start of every record. d1a from jonpol01/d1a#261 on (merged as b4b0f989) puts it back (bos_token_id 2; d1a_config.json records leading_ids: [2]). If you encode requests yourself, prepend id 2.

JohnP1/d1a-e2b-it v0.6 for Apple Silicon:

the adapter merged into Gemma 4 E2B-it; linear layers and token embeddings 8-bit (group 64), per-layer embeddings 4-bit;
pointer head fp32 (head.safetensors);
calibration temperature 1.41 in d1a_config.json;
3.5 GB on disk, plus 0.9 GB of photo, voice and video encoders in media/ (unchanged from the -it base).

Mac check on an Apple M1 Max, 28 real requests of up to 512 tokens (42 questions), against the fp32 PyTorch rebuild of this export: p50 199 ms, p95 294 ms per request (28 requests), 9 s to load; 0 of 42 answers change (0 of 40 clear-cut ones, where the reference's top answer leads by more than 0.05); largest probability change 0.015.

The scores below were measured on this build.

Release note. The bar registered for this run before training (jonpol01/d1a#227, Bar B, "good enough for apps") failed on two lines: severity −4.3 (it needed −3 or better) and media agreement 21/28 (it needed 25/28). It is published by the owner's decision of 2026-10-12 for an iOS proof of concept, as a separate repository and not as E2B's model.

What's inside

A LoRA adapter (rank 16 on q, k, v, o, gate, up, down) and a small pointer head (256-dimensional) on google/gemma-4-E2B-it at revision 3e22461f, merged and quantized here. Training: stage 1 is recipes/d1a-e2b.yaml (decision-v7, 2 epochs) from the -it trunk; stage 2 is E2B v0.6's run on its mix of 23,427 records (PR labels, hard, developer-tool and long-document decisions, JGLUE, routing, decision-v7 replay), 1 epoch from stage 1. One NVIDIA L4 on Hugging Face Jobs, about $3.51. Details: the PyTorch repo's card and provenance.json.

Calibration: a single temperature, T = 1.41 (90% interval [1.35, 1.48]), fitted on pooled rows of the decision-v7 calibration split and the PR-labelling development set. Calibration error is 0.053 before and 0.020 after, out of fold.

Results

Pull requests newer than all training

487 hermes-agent pull requests created after every PR in the training mix, held back before training. All models are MLX 8-bit at their served temperatures (E2B-it 1.41, E2B v0.6 1.45), with the serving context. Differences are paired over PRs, with 95% bootstrap intervals. E4B v0.6 is shown for reference:

Question	Majority answer	E2B-it v0.6	E2B v0.6	E4B v0.6	E2B-it − E2B v0.6
Severity (P0–P4)	67.8%	72.9%	77.2%	79.1%	−4.3 [−7.8, −1.0]
Change type	51.5%	87.3%	87.5%	90.6%	−0.2 [−3.1, +2.7]
P0 recall (7 PRs)		2/7	0/7	5/7
P1 recall (23 PRs)		1/23	0/23	8/23
Security type recall (14 PRs)		11/14	12/14	9/14

Calibration error on these 974 questions: 0.021.

Every other suite

Not measured yet. D1A's card-suite gate (transfer-v4, hard-v1, devtools-v1, documents-v1, decision-v7, JGLUE) has not run on this build. Do not assume E2B v0.6's numbers carry over.

Photos, voice and video

With the media extra, d1a.serving.serve answers questions about a photo, a voice clip or a video with this same model through Gemma 4 E2B-it's own vision and audio encoders in media/, unchanged by the adapter (POST /v1/systemone/media). No media in training (zero-shot), measured on only the playground's 14 photo, voice and video demo requests. It answers all 14:

21 of 28 answers match E4B v0.6 (E2B v0.6: 25 of 28). It misses damage E4B sees (the crushed-truck photo and video, the wet-parcel video), calls two voice clips urgent (the English directions and the Japanese not-at-home message), and places the damaged-door and intact-mailbox photos "on a sidewalk".
19 of 28 match E2B v0.6. Unlike E2B v0.6, it calls the intact locker and mailbox photos intact.

The questions it was trained for

PR labelling works best with these questions, worded exactly so, over a document of the form title / author / stats / body / files (one - status path +added/-deleted line per file):

type (choice): "Primary change type from files and body, not the title prefix." Options: type/bug, type/docs, type/feature, type/perf, type/refactor, type/security, type/test.
blast (choice): "How far a mistake in this PR spreads in production." Options: review:blast-contained (one module), -moderate (one subsystem), -broad (shared helper or config), -massive (auth, permissions, or all paths).
sev (choice): "How serious the problem this PR addresses is — not the risk of merging the diff as-is." Options P0 to P4.

The PR-labeler recipe has the exact wording, the document builder and the training and scoring scripts.

Run it

Needs d1a with jonpol01/d1a#261 (the <bos> above):

pip install "d1a[serve,mlx,media] @ git+https://github.com/jonpol01/d1a"
python -m d1a.serving.serve --run JohnP1/d1a-e2b-it-mlx-q8@v0.6 --port 8009

or in-process: from d1a import D1A; D1A.load("JohnP1/d1a-e2b-it-mlx-q8@v0.6").decide(state, questions). The PyTorch checkpoint is JohnP1/d1a-e2b-it v0.6; a 4-bit build is JohnP1/d1a-e2b-it-mlx-q4.

Limits

Experimental, and below E2B v0.6 where measured: severity −4.3 points on newer PRs, media 21/28 against 25/28 agreement with E4B v0.6. Its pre-registered bar failed (above).
Not a PR labeller. E4B v0.6 remains the PR labeler. On the 487 newer PRs this build finds 3 of 30 P0/P1 PRs (E4B v0.6: 13 of 30).
Most suites are not measured: hard, developer-tool and long-document decisions, held-out transfer, general decisions (decision-v7) and Japanese (JGLUE).
Gate tool calls through D1A's fail-safe thresholds (d1a.agents.presets.advise: deny at p ≥ 0.5, allow only at p ≥ 0.8, otherwise ask), never through the raw top answer. This build's demo answers have not been read yet.
<bos>: requests encoded without the leading id 2 are out of distribution (see above).
Media: zero-shot, measured on 14 demo requests only. It misses damage on 3 of them (above).
One temperature for every question, and no use-case temperatures.
The recent PR labels come from one large open-source project and a few days of it. If its conventions move again, so will these numbers.
Questions are in English. The documents can be English or Japanese (trained on both; Japanese not measured on this build).

Versions

Tag	What it is
`v0.6`	E2B v0.6's recipe on the Gemma 4 E2B-it trunk, MLX 8-bit with `media/`; experimental, Bar B of #227 failed

License and data

Apache-2.0. Base model: Gemma 4 by Google (Apache-2.0). Code: github.com/jonpol01/d1a, built on Kev by Jared Palmer (Apache-2.0). Japanese decision data derived from JGLUE by Yahoo Japan Corporation and Waseda University (CC BY-SA 4.0). Pull-request data from NousResearch/hermes-agent (MIT); the labelled PR dataset itself, including the recent PRs, is private.