D1A-E2B-it QAT · MLX 4-bit

This repository mirrors JohnP1/d1a-e2b-it-qat-mlx-q4 (revision 5bff9582) byte for byte (the same files, checked by LFS sha256); only this card differs. Its tags match the source's.

What it does: decisions (/v1/systemone), chat (OpenAI-format /v1/chat/completions, streaming included), and photos, voice clips and video (/v1/systemone/media). One model on Apple Silicon (MLX), 2.7 GB plus 1.0 GB of encoders.

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.

Versions

Only the adapter, the pointer head and the calibration temperature differ between the two; the trunk, the encoders and every file's shape are the same. The v0.6 card has v0.6's scorecard and the QAT retrain's story.

What v0.7 is better at, and what it is not

Measured on an Apple M1 Max, both versions as served, Δ in points with paired 95% intervals.

Better: dialogue-act decisions in English and Japanese (498 held-out questions written for D1A, by a different writer than the training questions, in ten domains the training rows did not use, such as a video call, a library, gardening and a museum guide; the closest pair is hotel services here and hotel booking in training):

The questions: did the user agree to what the assistant asked (yes / no / unclear: a counter-question, a hedge, a correction)? Is this only thanks or an acknowledgement, or is there a request after it? Which command did the user mean, or were they just talking? Which listed item do they mean?

Not better: intent classification. On MASSIVE 1.1 (60 intents, English and Japanese, with a "none of these" option): scenarios it trained on 70.8 → 69.7 (−1.2 [−3.7, +1.5]); scenarios held out of training 78.0 → 80.0 (+2.0 [−0.3, +4.3]). v0.7 picks "none of these" a little more often when a real intent fits (37 times against 30 in 600).

Everything else (every D1A suite: scripts/quality_gate.py --all-suites, scripts/decide.py):

• pooled over 22 suites: −0.38 [−0.87, +0.09]; one significant win (devtools-v1 +1.7 [+0.3, +3.0]), no significant loss;
• calibration slightly worse on some suites: the refitted temperature is 1.91 (v0.6: 1.55), and expected calibration error rose on typesafe (0.061 → 0.100), JGLUE test (0.027 → 0.048), WANLI (0.175 → 0.205, 0.187 → 0.215) and the hand-checked PRs (0.104 → 0.149); it fell on documents-v1 (0.045 → 0.026) and night2 dates (0.042 → 0.022);
• the demos: 3 answers changed across the 145 demo requests. One voice clip's intent is now right ("the customer is not home", which v0.6 read as "report a damaged parcel"); two urgent flags that were near ties flipped to urgent;
• speed: unchanged. v0.7 has v0.6's shape and does the same work per answer. decide.py's latency veto (V3) did fire (v0.7 against v0.6 1.018, line 1.016). A follow-up measurement put a byte-identical copy of v0.6 in the same slot and got 1.032, so the gap comes from serving two different files, not from v0.7. That audit's own pre-set rule still kept the veto, and v0.7 is released by the owner's decision of 2026-10-30.

What's inside

• The trunk: Gemma 4 E2B-it QAT (google/gemma-4-E2B-it-qat-w4a16-ct, revision 971342c0), its 276 int4 linear layers stored bit for bit as MLX 4-bit (the same packed words and scales, dequantization error 0.0); token and per-layer embeddings 4-bit in groups of 64. 2.7 GB.
• The D1A adapter, unmerged (adapter/, LoRA rank 16, 97 MB). With the adapter off the model is Gemma 4 E2B-it QAT itself, so the same weights also chat.
• The pointer head (head.safetensors).
• Calibration temperature 1.91 in d1a_config.json, fitted on this build's own answers: 3,178 rows from the decision-v7 calibration split and the PR-label development set, v0.6's recipe. Expected calibration error 0.084 → 0.030 out of fold. Both sets come from corpora the model trained on, which the fit records.
• media/: Gemma 4's vision and audio encoders from the QAT checkpoint (bf16, 1.0 GB), unchanged from v0.6.

Training (v0.7)

• warm start from v0.6's adapter, on the same exact QAT trunk;
• one epoch of 7,008 records: 2,100 from MASSIVE 1.1's training split (English and Japanese, 20 utterances per intent and language; every utterance for the four rarest intents), 1,248 dialogue-act questions written for D1A (English and Japanese; private), 60 amount-confirmation questions, and 3,600 replayed from every earlier D1A training source (PR labels in English and Japanese, blast radius, hard, devtools, documents, JGLUE, routing, decision-v7), so no earlier skill went untrained;
• the evaluation questions are kept apart from training: different MASSIVE ids, two scenarios held out entirely (transport, email), and dialogue-act domains the training rows did not use, screened for near-duplicates;
• learning rate 2e-5, batch 8, 876 steps, 3.2 hours on an Apple M1 Max (MPS);
• exactness: 276 of 276 quantized linear layers identical to the QAT checkpoint (words, scales and biases).

Run it

pip install "d1a[serve,mlx,media] @ git+https://github.com/jonpol01/d1a"
python -m d1a.serving.serve --run mlx-community/d1a-e2b-it-qat-4bit --port 8009

v0.6 instead: --run mlx-community/d1a-e2b-it-qat-4bit@v0.6.

A decision:

curl -s localhost:8009/v1/systemone -H 'content-type: application/json' -d '{
  "state": "Assistant: Shall I cancel the 3 pm meeting?\nUser: is that the one with the client?",
  "questions": {"answer": {"type": "choice", "instructions": "Did the user agree to the assistant'"'"'s question?",
                           "criteria": {"yes": "Agrees: go ahead", "no": "Refuses, wants to wait, or corrects it",
                                        "unclear": "Neither: a counter-question, a hedge, or something unrelated"}}}
}'

It answers unclear (0.62), a counter-question; "not yet, I need to check first" answers no (0.96) and "sure, why not" yes (0.99). About 70 ms per decision on an Apple M1 Max.

A photo, voice clip or video: the same request with a media field. type is image (JPEG, PNG, WebP), audio (WAV, FLAC or OGG, up to 30 s) or video (MP4, MOV or WebM, up to 32 MB; 16 frames sampled evenly, the sound track not used):

curl -s localhost:8009/v1/systemone/media -H 'content-type: application/json' -d '{
  "media": {"type": "image", "data": "'"$(base64 -i parcel.jpg)"'"},
  "questions": {"damaged": {"type": "noul", "instructions": "Is the parcel damaged?"}}
}'

Chat (the adapter is switched off for it, so this is Gemma 4 E2B-it QAT itself):

curl -s localhost:8009/v1/chat/completions -H 'content-type: application/json' -d '{
  "messages": [{"role": "user", "content": "Name three things to check when a parcel arrives damaged."}]
}'

or in-process: from d1a import D1A; D1A.load("mlx-community/d1a-e2b-it-qat-4bit").decide(state, questions).

Limits

• Act on a "yes" only above a confidence threshold. In dialogue-act tests v0.7 still reads some refusals and hedges as agreement; an app that acts on the answer should confirm when the top probability is below about 0.6.
• Intent classification did not improve (above), and calibration is a little worse on some suites. For task types far from D1A's own, the v0.6 card's transfer-v4 note still applies.
• Photos, voice and video were checked on the demo requests only, zero-shot as on every D1A-E2B-it build.
• <bos>: the model was trained with id 2 at the start of every record; d1a adds it (leading_ids: [2]). An app that encodes requests itself must prepend it.
• Questions are in English; the documents can be English or Japanese.
• Everything on the D1A-E2B-it card applies: experimental, and agent tool calls go through D1A's fail-safe thresholds (d1a.agents.presets.advise).

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). Intent data from MASSIVE 1.1 by Amazon (FitzGerald et al., 2022; CC BY 4.0), English and Japanese. 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.