README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-4B
library_name: anyjev
pipeline_tag: text-generation
language: [en]
version: v1
tags: [anyjev, tacit, decision-model, typed-decisions, churn, customer-success, slack, self-distillation, lora, v1, qwen3-4b]
datasets: [workmagicio/slack-churn]
metrics: [accuracy, f1, expected_calibration_error]
---

# workmagic-jev-decision

- 版本：v1
- 基座：`Qwen/Qwen3-4B`

用 AnyJev **自蒸馏**（self-distillation）得到的一次前向决策模型：把老师 `Decider(level=L0)` 在 **Slack 渠道流失（churn）** 决策上的答案蒸馏进单次前向读出；训练**无人工标签**。

## 用法
```python
from anyjev import Tacit
tacit = Tacit.from_pretrained("workmagicio/workmagic-jev-decision", adaptive=False)   # 可用 revision="v1" 固定版本
d = tacit.decide(
    state="...slack conversation...",
    question="...how likely is the customer to churn...",
    options=["low", "medium", "high"],
)
print(d["answer"], d["probs"])
```

## 训练
基座 `Qwen/Qwen3-4B`，采用 AnyJev readout 提示词格式，LoRA 自蒸馏（只在答案 token 上算 loss）。