D1A-E4B-it QAT · MLX 4-bit

This repository mirrors JohnP1/d1a-e4b-it-qat-mlx-q4 (revision b4a6a5d1) byte for byte (the same files, checked by LFS sha256); only this card differs.

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), 4.4 GB plus 0.9 GB of encoders (D1A-E4B-it v0.6 in 8-bit: 6.1 GB plus 0.9 GB).

Scorecard (2026-10-28): level with D1A-E4B-it v0.6 overall, stronger on PR labels and routing, weaker on one external suite (below). Released 2026-10-27 by the owner's decision, before its gates finished; both gates were set before training and have since run on every D1A suite, the demos and the PR labeler's data.

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 build sits exactly on Google's quantization-aware-trained int4 weights, google/gemma-4-E4B-it-qat-w4a16-ct (revision 6cd26aaa). All 343 of its int4 linear layers are stored here bit for bit as MLX 4-bit: the same packed words and scales, dequantization error 0.0. Nothing is re-quantized from bf16. The D1A adapter was retrained on this exact trunk, so the 4-bit model is the model that was trained, not a rounded copy of it.

Why it was retrained:

The QAT weights differ from the bf16 ones the v0.6 adapter was trained on. On E2B, the v0.6 adapter moved onto its QAT trunk unchanged lost accuracy and changed 9.6% of its clear-cut answers (D1A-E2B-it QAT).
D1A-E4B-it v0.6 was trained with PyTorch on an Apple GPU, where a PyTorch bug read Gemma 4's per-layer embeddings as zeros for every token id from 199,729: every punctuation and digit token, 28.5% of its training tokens (jonpol01/d1a#353, pytorch/pytorch#188756). This retrain reads them correctly (d1a with jonpol01/d1a#354).

Scorecard

Against D1A-E4B-it v0.6 (MLX 8-bit): level overall, pooled −0.16 points [−0.88, +0.51] over 21 lines, with no veto from D1A's release rules. Against v0.6's adapter moved unchanged onto this same QAT trunk: +1.19 [+0.50, +1.87], 6 significant wins and no significant loss, so the retrain is what makes the 4-bit trunk work.

Stronger than v0.6: PR labels on the test set +1.6, PR severity with the per-class offsets +6.0, PR blast radius +7.2, routing +1.7 and +1.5.
Weaker than v0.6: the external typesafe suite, −7.8 (jonpol01/d1a#361). It comes from the 4-bit trunk, not the retrain: v0.6's own adapter on this trunk loses nearly as much (−6.9). Calibration is looser on transfer-v4 (ECE 0.081 → 0.105). With the severity offsets on, it catches fewer P1 pull requests (29 of 47 against 39 of 47) at a higher precision (0.30 against 0.22).
Speed: about the same per decision as v0.6 in 8-bit (ratio 1.01, inside D1A's latency limit), at 5.3 GB instead of 7.0 GB.
Photos, voice and video: of the 28 answers to the 14 demo requests, 26 are the same as v0.6's. The two that differ: an intact parcel in a mailbox read as "in a delivery locker", and a voice clip about a damaged parcel marked urgent.

Every suite, accuracy in %, Δ in points with paired 95% bootstrap intervals (bold where the interval excludes 0); measured with scripts/quality_gate.py --all-suites and scripts/decide.py on an Apple M1 Max. ECE and the method: the full scorecard.

Suite	n	This build	Δ vs D1A-E4B-it v0.6	Δ vs v0.6's adapter on this trunk
devtools-v1 (card)	1,074	74.6	+0.2 [-1.1, +1.5]	+0.3 [-1.0, +1.6]
documents-v1 (card)	920	88.9	+0.3 [-1.1, +1.7]	+1.1 [-0.2, +2.4]
external semif-v1	144	88.9	-2.1 [-5.6, +0.7]	-1.4 [-4.9, +2.1]
external typesafe-v1	102	79.4	-7.8 [-11.7, -3.6]	-1.0 [-5.8, +4.3]
external wanli-v1	256	72.3	-1.6 [-3.9, +0.8]	+0.4 [-2.3, +3.1]
external wanli-v2	1,002	72.1	-0.5 [-1.9, +0.9]	+1.6 [+0.1, +3.1]
hard-v1 (card)	1,083	76.8	+1.2 [-0.4, +2.9]	+3.4 [+1.7, +5.4]
ja-jglue development	1,500	80.7	-0.1 [-1.3, +1.1]	+1.0 [-0.2, +2.3]
ja-jglue test	1,500	79.7	+1.2 [+0.0, +2.4]	+1.7 [+0.4, +3.0]
night2 assertion	1,020	94.1	+0.9 [-0.3, +2.0]	+1.7 [+0.7, +2.7]
night2 dates	900	90.0	-1.2 [-2.7, +0.2]	+0.9 [-0.2, +2.0]
night2 unknowable	525	67.2	+0.6 [-0.8, +1.9]	+0.6 [-1.0, +2.3]
pr-labels development	2,030	80.5	+0.9 [-0.1, +1.9]	+1.7 [+0.4, +3.0]
pr-labels development-ja	198	78.8	+3.5 [-0.5, +7.2]	-0.5 [-5.0, +4.1]
pr-labels real17	51	64.7	-3.9 [-11.8, +3.9]	+0.0 [-7.8, +7.8]
pr-labels test	2,045	81.0	+1.6 [+0.5, +2.7]	+4.4 [+3.0, +5.9]
pr-labels test-ja	193	76.2	-0.5 [-2.7, +1.6]	+3.1 [-1.0, +7.4]
routing factory-development	640	95.2	+1.7 [+0.3, +3.0]	+0.6 [-0.9, +2.2]
routing generic-development	270	98.9	+1.5 [+0.4, +3.0]	+1.1 [+0.0, +2.6]
routing handlabelled-45	45	100.0	+0.0 [+0.0, +0.0]	+0.0 [+0.0, +0.0]
transfer-v4 (card)	656	76.1	-1.2 [-3.4, +0.8]	+0.0 [-1.8, +1.7]
decision-v7 (card)	1,264	86.0	+0.2 [-0.8, +1.3]	+0.7 [-0.2, +1.6]
PR severity, with offsets (test)	953	75.3	+6.0 [+3.8, +8.3]	+8.6 [+6.1, +11.0]
PR blast radius (test)	139	64.0	+7.2 [+1.4, +13.7]	+7.9 [+0.7, +14.4]
PR labeler replay (recent PRs)	193	72.0	+1.0 [-4.2, +6.2]	+2.1 [-3.1, +7.2]
owner's PR labels	109	71.6	+0.9 [-4.4, +6.5]	+4.6 [+0.0, +9.3]

What's inside

The trunk: Gemma 4 E4B-it QAT, its 343 int4 linear layers exactly as Google trained them (symmetric, groups of 32); token and per-layer embeddings 4-bit in groups of 64 (QAT fixes only the Linears' groups of 32). 4.4 GB.
The D1A adapter, unmerged (adapter/, LoRA rank 16, 133 MB). With the adapter off the model is Gemma 4 E4B-it QAT itself, so the same weights also chat.
The pointer head (head.safetensors).
Calibration temperature 1.66 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, the same recipe as v0.6. Expected calibration error 0.069 → 0.016 out of fold; 90% interval of the temperature 1.59–1.74. Both sets come from corpora the model trained on, which the fit records.
Routing temperature 1.41 (use_case_temperatures.routing, added 2026-11-08, d1a#382): requests sent with "use_case": "routing" are read at 1.41; every other request keeps 1.66, and no answer's top choice changes. It was fitted as v0.6.1's routing temperature was, on held-out routing rows (generic-development + handlabelled-45, 315 rows). That fit set is uninformative (accuracy 0.99), so 1.41 is the method's point estimate, confirmed directionally out of sample on factory-development (expected calibration error 0.049 at 1.66 → 0.035 at 1.41, change −0.014 with 95% interval −0.022 to −0.006; same corpus as training). It is not claimed optimal; 1.0 scored 0.016 there and is a sensitivity line.
media/: Gemma 4's vision and audio encoders from the QAT checkpoint itself (bf16, 0.9 GB).
media_pad_row.safetensors (jonpol01/d1a#394; tag t1-media-pad-row): the pad token's per-layer embedding row as the QAT checkpoint stores it (10,752 bf16 values, 22 KB), named media_pad_row in d1a_config.json. Photos, voice and video read it at their media positions instead of the 4-bit table's row, whose error added up over hundreds of positions. On the 14 demo media requests the answers move by up to 0.108 (mean 0.023) and one top answer changes (an intact parcel in a mailbox: delivery locker → mailbox); text answers do not change. There is no full-precision media reference for E4B; on D1A-E2B-it QAT the same fix takes the largest change against full precision from 0.140 to 0.085. D1A reads it from jonpol01/d1a#394 on (main after 0.7.0); an older D1A ignores it, and the routing-t1.41 tag predates it.

Training

D1A-E4B-it v0.6's recipe, continued on the QAT trunk (jonpol01/d1a#341):

warm start from JohnP1/d1a-e4b-it v0.6;
one epoch of v0.6's all-skills mix: 23,427 records from 11 sources (PR labels in English and Japanese, blast radius, recent PRs, hard, devtools, documents, JGLUE, routing, decision-v7);
the trunk frozen at the QAT checkpoint's values, dequantized exactly (scale × integer in fp32), the embedding tables in bf16 as the checkpoint stores them (--embedding_dtype bf16, jonpol01/d1a#355); learning rate 2e-5, batch 8;
2,929 steps, 36.2 hours on an Apple M1 Max (MPS).

Other checks

Exactness: 343 of 343 quantized linear layers identical to the QAT checkpoint (words, scales and biases).
Serving (d1a.serving.serve, Apple M1 Max): the 145 playground demo requests (the 14 photo, voice and video requests included) and 92 PR-labeler states all answered; the example request below answers billing (0.40, shipping 0.37); with the adapter off, chat answers.

Run it

Needs d1a with jonpol01/d1a#261:

pip install "d1a[serve,mlx,media] @ git+https://github.com/jonpol01/d1a"
python -m d1a.serving.serve --run JohnP1/d1a-e4b-it-qat-mlx-q4 --port 8009

A decision:

curl -s localhost:8009/v1/systemone -H 'content-type: application/json' -d '{
  "state": "Shoes arrived two weeks late and in the wrong size. Also I see two charges on my card.",
  "questions": {"team": {"type": "choice", "instructions": "Which team should handle this?",
                         "criteria": {"returns": null, "shipping": null, "billing": null}}}
}'

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 E4B-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("JohnP1/d1a-e4b-it-qat-mlx-q4").decide(state, questions).

Limits

The external typesafe suite: −7.8 points against D1A-E4B-it v0.6 (above, jonpol01/d1a#361). Pin JohnP1/d1a-e4b-it-mlx-q8 v0.6 where answers must match v0.6's.
Overconfident on transfer-v4 (eight public task sources held out of training): mean confidence is 0.10 above accuracy (ECE 0.105; D1A-E4B-it v0.6: 0.08 above, ECE 0.081), from about 0 (sciq) to +0.25 (paws, emotion) by source. Of answers given at ≥ 0.9, 10.7% are wrong (v0.6: 8.8%). One global temperature does not fix it: the temperature this suite alone wants (2.9, against the served 1.66) makes 15 of the 21 other suites worse (jonpol01/d1a#361).
Photos, voice and video are measured on the 14 demo requests only, zero-shot as on every D1A-E4B-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-E4B-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). 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.