| --- |
| license: apache-2.0 |
| base_model: google/gemma-4-E2B |
| base_model_relation: adapter |
| library_name: peft |
| pipeline_tag: text-classification |
| tags: [d1a, decision-model, calibration, lora, gemma4, typesafe, system-one] |
| --- |
| |
| # D1A-E2B |
|
|
| **D1A** is a small open decision model in the Jev style: one document and a set of typed questions in, a calibrated probability for every option out, in one forward pass, with no text generation. It speaks the same System One API as Jev, so the TypeSafe SDK works against it unchanged. |
|
|
| This checkpoint is a LoRA adapter plus a pointer head on `google/gemma-4-E2B` (revision `d29ff6b4`), trained for two epochs and calibrated with a temperature of 1.52. |
|
|
| | | D1A-E2B (this checkpoint, 2 epochs) | v0.1 (1 epoch) | Jev (TypeSafe, hosted) | |
| |---|---|---|---| |
| | Accuracy, trained sources (dev) | 0.824 | 0.794 | 0.845 | |
| | Accuracy, new sources (dev) | 0.602 | 0.569 | 0.857 | |
| | Log loss (dev) | 0.499 | 0.522 | | |
| | Calibration error, ECE (dev) | 0.057 | 0.040 | | |
| | Where it runs | your machine (Apple GPU / NVIDIA), free, private | | TypeSafe's cloud API | |
|
|
| Measured on the same frozen evaluation sets (development partitions; the locked test was not read). Jev is more accurate today. The second epoch gained 3 points on both sets at a slightly higher calibration error. Tag `v0.1-1epoch` keeps the previous version; the Apple Silicon build [JohnP1/d1a-e2b-mlx-q8](https://huggingface.co/JohnP1/d1a-e2b-mlx-q8) is still made from v0.1. |
|
|
| - Playground (nine live use cases): https://github.com/jonpol01/d1a-playground |
| - Code: https://github.com/jonpol01/d1a |
|
|
| ## Serve it |
|
|
| ```bash |
| pip install "d1a[serve] @ git+https://github.com/jonpol01/d1a" |
| python -m d1a.serve --run JohnP1/d1a-e2b --port 8009 |
| ``` |
|
|
| ## License |
|
|
| Apache-2.0. Base model: Gemma 4 by Google (Apache-2.0). Training and serving code: github.com/jonpol01/d1a, built on Kev by Jared Palmer (Apache-2.0). |
|
|
