| license: apache-2.0 | |
| base_model: autotrust/GEV-26B-Decide | |
| library_name: transformers | |
| tags: | |
| - decision | |
| - systemone | |
| # autotrust/GEV-26B-Decide: decision files | |
| Chat template, settings and head of [autotrust/GEV-26B-Decide](https://huggingface.co/autotrust/GEV-26B-Decide) for `pipeline("decision")` | |
| ([transformers#49566](https://github.com/huggingface/transformers/pull/49566)). The backbone is google/gemma-4-26B-A4B-it with the System 1 LoRA of the original (`adapter_model.safetensors`), loaded by `from_pretrained`. | |
| ```python | |
| from transformers import pipeline | |
| # the LoRA at the root makes `pipeline` load its base model, so the decision files are passed explicitly | |
| decide = pipeline("decision", model="ArthurZ/GEV-26B-Decide", decision_config="ArthurZ/GEV-26B-Decide") | |
| decide({"state": "I was charged twice and want a refund.", "questions": {"refund": {"type": "noul", "instructions": "Is a refund requested?"}}}) | |
| ``` | |
| - `decision_template.jinja`: the chat template; `<|mark|>` marks the positions the head reads. | |
| - `decision_config.json`: temperatures and option order. | |
| - `decision_head/`: `JevDecisionHead`, converted from `head.safetensors` of the original. | |
| - `decision_test.json`: a request and the input ids that the original code builds for it, used by the tests. | |
| Derived from [autotrust/GEV-26B-Decide](https://huggingface.co/autotrust/GEV-26B-Decide) (Apache-2.0). | |