JohnP1 commited on
Commit
0e161cb
·
verified ·
1 Parent(s): 1b8cf90

D1A-E4B-it v0.6: README.md

Browse files
Files changed (1) hide show
  1. README.md +73 -0
README.md ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: google/gemma-4-E4B-it
4
+ library_name: mlx
5
+ pipeline_tag: text-classification
6
+ language: [en, ja]
7
+ tags: [d1a, decision-model, pull-requests, mlx, apple-silicon, quantized, 4-bit, on-device]
8
+ ---
9
+
10
+ <p align="center"><img src="https://huggingface.co/JohnP1/d1a-e4b-it-mlx-q4/resolve/main/d1a-avatar.png" width="200" alt="D1A avatar: an anime girl with silver-white and dark purple hair, round glasses and headphones on a round #D1A badge"></p>
11
+
12
+ # D1A-E4B-it v0.6 · MLX 4-bit
13
+
14
+ **A small decision model for Apple Silicon: give it a document and a few questions, and it picks the answers.** Each question
15
+ is yes/no, multiple choice or a score; D1A returns a calibrated probability for every option in one forward pass, with no
16
+ generated text, so every answer is a valid option.
17
+
18
+ This is the 4-bit MLX build of [JohnP1/d1a-e4b-it](https://huggingface.co/JohnP1/d1a-e4b-it) v0.6, with the adapter merged
19
+ into its weights. Scores, training and limits are on that card; how far this build moves from the served chat build is below.
20
+
21
+ ```python
22
+ from d1a import D1A # pip install "d1a[serve,mlx,media] @ git+https://github.com/jonpol01/d1a" (needs jonpol01/d1a#261)
23
+ m = D1A.load("JohnP1/d1a-e4b-it-mlx-q4@v0.6")
24
+ m.decide("The customer was charged twice for the same order and wants a refund today.",
25
+ {"team": {"type": "choice", "instructions": "Which team handles this?",
26
+ "criteria": {"billing": None, "shipping": None, "support": None}},
27
+ "angry": {"type": "noul", "instructions": "Is the customer angry?"}})
28
+ ```
29
+
30
+ **Not a chat model.** This build has the adapter merged into its weights, so it answers D1A decisions only; text generated
31
+ from it would come from a changed backbone, not Gemma 4 E4B-it. To chat with the same model that decides, build an unmerged
32
+ folder from the adapter: see [Enable chat](https://huggingface.co/JohnP1/d1a-e4b-it#enable-chat).
33
+
34
+ **`<bos>`: encode with it.** The Gemma 4 E4B-it tokenizer adds no `<bos>`, but this model was trained with one at the start
35
+ of every record (`d1a_config.json` records `leading_ids: [2]`). If you encode requests yourself, prepend id 2.
36
+
37
+ ## What's inside
38
+
39
+ - the LoRA adapter merged into Gemma 4 E4B-it (revision `ee0ef602`); linear layers and token embeddings 4-bit (group 64), per-layer embeddings 4-bit;
40
+ - pointer head fp32 (`head.safetensors`);
41
+ - calibration temperature 1.55 in `d1a_config.json`, the run's, not refitted for this build; no use-case temperatures;
42
+ - 3.9 GB on disk for the text model, plus 0.9 GB of photo, voice and video encoders in `media/` (unchanged from the -it base).
43
+
44
+ ## Against the served build
45
+
46
+ Same code, this build against the unmerged 8-bit chat build the Mac mini serves (`quality_gate.py`, Apple M1 Max): 145 demo requests (202 questions) and 92 PR-labeler states (276 questions).
47
+ - **9 of 202** demo answers and **33 of 276** labeler answers (12%) change; largest probability change 0.35;
48
+ - model time 0.83× the chat build's.
49
+
50
+ **This 4-bit build has not been scored on the evaluation suites.** It moves noticeably more than the 8-bit build; prefer [-mlx-q8](https://huggingface.co/JohnP1/d1a-e4b-it-mlx-q8) unless memory decides.
51
+
52
+ ## Run it
53
+
54
+ ```bash
55
+ pip install "d1a[serve,mlx,media] @ git+https://github.com/jonpol01/d1a"
56
+ python -m d1a.serving.serve --run JohnP1/d1a-e4b-it-mlx-q4@v0.6 --port 8009
57
+ ```
58
+
59
+ ## Limits
60
+
61
+ - Everything on the [D1A-E4B-it card](https://huggingface.co/JohnP1/d1a-e4b-it#limits) applies.
62
+ - **Not scored on the suites at 4 bits**: 12% of PR-labeler answers differ from the served build; expect the card's numbers to move.
63
+ - **`<bos>`**: requests encoded without the leading id 2 are out of distribution.
64
+
65
+ ## Versions
66
+
67
+ | Tag | What it is |
68
+ |---|---|
69
+ | **`v0.6`** | **MLX 4-bit build of D1A-E4B-it v0.6 (adapter merged)** |
70
+
71
+ ## License and data
72
+
73
+ Apache-2.0. Base model: Gemma 4 by Google (Apache-2.0). Code: [github.com/jonpol01/d1a](https://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.