| --- |
| license: apache-2.0 |
| tags: |
| - yev |
| - decision-model |
| --- |
| |
| # yev |
|
|
| The client for [Yev-Decision-1](https://huggingface.co/HallD/Yev-Decision-1), a local **System One** |
| decision model. Give it a state, which can be text, images or both, plus named questions. It returns |
| a typed probability distribution for each one: a **Choice** between options, a **Score** on an |
| ordered scale, or a **Noul** (yes or no). It generates no text and nothing leaves the machine. |
|
|
| ## Install |
|
|
| ```sh |
| uv add "yev[inference] @ git+https://huggingface.co/HallD/yev" |
| ``` |
|
|
| `pip install "yev[inference] @ git+https://huggingface.co/HallD/yev"` works the same way. Python 3.13 |
| or later. |
|
|
| | Extra | Brings | For | |
| | --- | --- | --- | |
| | *(none)* | pydantic, huggingface-hub, numpy | Question and answer types only | |
| | `inference` | torch, transformers, peft, accelerate, pillow | Answering questions with a model | |
| | `quantized` | `inference` plus bitsandbytes | `quantize4bit=True` on a CUDA GPU | |
|
|
| ## Use |
|
|
| ```python |
| from yev import Choice, Image, Noul, Score, YevClient |
| |
| with YevClient() as client: # YEV_DEFAULT_MODEL, else HallD/Yev-Decision-1 |
| response = client.system_one( |
| state="I was charged twice. Please fix this ASAP.", |
| questions={ |
| "billing": Noul(instructions="Is this ticket about billing?"), |
| "tone": Choice(instructions="What is the customer's tone?", |
| criteria={"calm": None, "frustrated": None, "angry": None}), |
| "urgency": Score(instructions="How urgent is this ticket?", |
| criteria=["can wait", "this week", "today"]), |
| }, |
| ) |
| print(response.nouls["billing"].noul, response.choices["tone"].probabilities) |
| ``` |
|
|
| The first call downloads the model and its backbone, |
| [`Qwen/Qwen3.5-4B-Base`](https://huggingface.co/Qwen/Qwen3.5-4B-Base) at a pinned revision, into the |
| Hugging Face cache. After that it runs offline (`HF_HUB_OFFLINE=1`). The backbone runs in bfloat16; |
| its weights and the adapter take about 10.4 GB. |
|
|
| The call shape follows the TypeSafe SDK (0.7.5): `YevClient` and `AsyncYevClient` take its |
| `system_one(state, questions, model=, assets=, response_model=)`, its question objects (TypeSafe's |
| own are accepted), and return its answer and response types. It runs in-process, so there are no |
| API keys, retries or timeouts; Box questions are refused, and a Score takes at most 10 levels. |
|
|
| `yev --export HallD/Yev-Decision-1 --state "I was charged twice." --noul billing "Is this ticket about billing?"` |
| asks from the command line. |
|
|
| ## JevBench |
|
|
| `examples/jevbench_adapter.py` is an adapter for the |
| [JevBench](https://github.com/fstandhartinger/jevbench) harness: it sends the harness's own question |
| to `system_one` and returns the probabilities, with the input tokens the model read as `usage` |
| (output tokens are always 0). |
|
|
