aimeigaoshou/agent-jev

AgentJev-0.6B
State in. A distribution out. Nothing decoded.

This is the coding-completion checkpoint. It reads a task, the code, and the public checks, then returns a distribution over whether the work is actually done. It does not write the decision one token at a time.

The run stays on Qwen3-0.6B. Supervised training and RLCD use executed coding pairs: the same public checks pass on both sides, and the label comes from the checks that were not shown. The development set chose RLCD step 12.

What it is good at

Check	Result
Coding completion, accuracy	57.8% · 606/1048
Coding completion, recall	57.8%
Coding completion, AUROC	0.589
Correct implementation ranked ahead of the broken one	348/524
Invoice processing	87.2%

Coding accuracy moved from 51.7% to 57.8%. Recall moved from 35.7% to 57.8%.

Use it

The serving path is a typed decision, not a chat completion. Send a state and a boolean question. The answer is the probability that the implementation meets the stated requirements.

from safetensors.torch import load_file
weights = load_file("model.safetensors")

Weights are FP32. The checkpoint is the RLCD step selected on the frozen development set. The test split was not used to choose it.

---

AgentJev-0.6B
状态进去，分布出来，不逐词生成结论。

这一版是代码完成判断。它读任务、代码和已经公开的检查，然后给出「这件工作是否真正做完」的分布。

底座仍是 Qwen3-0.6B。监督训练和 RLCD 用的是执行过的代码对照：两边都通过同一组公开检查，标签来自没有展示出来的检查。开发集选中 RLCD 第 12 步。

这一版的成绩

项目	结果
代码完成，准确率	57.8% · 606/1048
代码完成，召回	57.8%
代码完成，AUROC	0.589
正确实现排在有缺陷实现前面	348/524
发票处理	87.2%

代码题准确率从 51.7% 提到 57.8%，召回从 35.7% 提到 57.8%。

调用方式是带类型的决策，不是聊天补全。送入状态和一道是否完成的问题，返回实现是否满足需求的概率。权重是 FP32。测试划分没有参与选模型。