Model card

README.md

@@ -3,31 +3,62 @@ license: mit
 library_name: transformers.js
 pipeline_tag: feature-extraction
 base_model: BAAI/bge-small-en-v1.5
-language: en
+language:
+- en
 tags:
 - decidekit
 - system-one
 - text-classification
+- zero-shot-classification
+- calibration
 - onnx
+- on-device
+thumbnail: https://huggingface.co/debuffjp/decide-small/resolve/main/assets/banner.png
 ---
 
+![Decide Small](https://huggingface.co/debuffjp/decide-small/resolve/main/assets/banner.png)
+
 # Decide Small
 
-The sentence encoder behind Decide Small, a System One model that picks one of your labels for a piece of text on the user's device. English. 34 MB at 8-bit. [decidekit](https://debuff.jp/decide/) adds the trained head and calibrated confidence.
+Decide reads a short piece of text and picks one of your labels, on the user's device, with a confidence you can act on: answers at 0.9 are right about 90% of the time. It's the default model in [decidekit](https://debuff.jp/decide/).
+
+[Docs](https://debuff.jp/decide/docs/) · [Try it in your browser](https://debuff.jp/decide/playground/) · [GitHub](https://github.com/debuffjp/decidekit) · [npm](https://www.npmjs.com/package/decidekit)
+
+## Use it
+
+```bash
+npm install decidekit
+```
 
 ```ts
 import { createDecider } from "decidekit";
 
-const d = await createDecider();
-const { choice, confidence } = await d.decide("my card still hasn't arrived", ["card_arrival", "refund"]);
+const d = await createDecider(); // loads this model
+const { choice, confidence } = await d.decide("my card still hasn't arrived", ["card_arrival", "lost_card", "refund"]);
 ```
 
-- BANKING77, 77 labels, all training data: 90.7%
-- CLINC150, 150 labels, all training data: 94.4%
-- A decision in Chrome (worker, median): 7.7 ms
+The files download from this repo on first use, then stay in the browser's cache. To serve them yourself, copy this repo's files to your own host and pass `modelHost` to `createDecider()`.
+
+## Numbers
+
+| | |
+|---|---|
+| BANKING77, 77 labels, all training data | 90.7% |
+| CLINC150, 150 labels, all training data | 94.4% |
+| BANKING77 with 5 examples per label | 82.3% |
+| A decision in Chrome (worker, median) | 7.7 ms |
+| Download at 8-bit | 34 MB |
+
+Ryzen 7 260 laptop, Chrome 156. Results and scripts: [eval/results-bge-small.json](https://github.com/debuffjp/decidekit/blob/main/eval/results-bge-small.json).
+
+## Limits
+
+- English. Use [Decide Multilingual](https://huggingface.co/debuffjp/decide-multilingual) for other languages.
+- It reads what a text is about better than how it feels: 61% on tweet sentiment.
+- About a paragraph at a time. Split longer text.
 
-Measured on a Ryzen 7 260 laptop. Scripts and results: [eval/results-bge-small.json](https://github.com/debuffjp/decidekit/blob/main/eval/results-bge-small.json).
+## What's in this repo
 
-## Where the weights come from
+This repo holds the encoder, the part that turns text into vectors. decidekit adds the trained head and the calibrated confidence.
 
-These files are an unmodified copy of [Xenova/bge-small-en-v1.5](https://huggingface.co/Xenova/bge-small-en-v1.5) at revision `ea104da`, the ONNX conversion of [BAAI/bge-small-en-v1.5](https://huggingface.co/BAAI/bge-small-en-v1.5), which is MIT licensed. We did not train them. They are published here so decidekit has a stable, pinned source.
+The weights are an unmodified copy of [Xenova/bge-small-en-v1.5](https://huggingface.co/Xenova/bge-small-en-v1.5) at revision `ea104da`, the ONNX conversion of [BAAI/bge-small-en-v1.5](https://huggingface.co/BAAI/bge-small-en-v1.5). We did not train them. They are here so decidekit has a stable, pinned source. MIT licensed, like the original.