从选项logits取分布,别把聊天生成当单步分类
本页展示随包原生输入和验收参考。目标产品运行状态:未运行。表格与图片不代表产品已经生成的结果。
input.csv / practice-data.json
本课实际输入字段
| ticket | state | authorTeam | options |
|---|---|---|---|
| T01 | 付款连续失败,账号仍可登录。 | payments | |
| T02 | 忘记密码,无法登录。 | account | |
| T03 | 周日客服几点上班? | other | |
| {"payments": "支付与付款故障", "account": "登录与账号访问", "other": "其他咨询"} |
准备单token标签
reference.csv:作者验收映射,不是目标API回包
| labels | supportedBases | readout |
|---|---|---|
| {"A": "payments", "B": "account", "C": "other"} | ||
| ["granite-4.1-8b", "granite-4.1-30b"] | ||
| softmax over option-label logits, one forward pass |
应当保留的约束
| labels | supportedBases | readout |
|---|---|---|
| {"A": "payments", "B": "account", "C": "other"} | ||
| ["granite-4.1-8b", "granite-4.1-30b"] | ||
| softmax over option-label logits, one forward pass |
故意构造的错误样例
| base | error | choiceOptions | method |
|---|---|---|---|
| granite-4.1-3b | Decision adapter not offered | ||
| choice requires at least2 | ["only one"] | ||
| not official one-step readout | generate free explanation then parse |
先保存真实产品输出,再逐项对照。actualOutput / actualResult 在未执行时保持 not_run,不能把这里的期望值复制成已测结果。