107 lines
2.2 KiB
Markdown
107 lines
2.2 KiB
Markdown
# 核心模型抽取样板
|
||
|
||
version: 0.2
|
||
|
||
last_updated: 2026-06-16
|
||
|
||
status: draft
|
||
|
||
## Source Documents
|
||
|
||
原始 GPT 规划位于：
|
||
|
||
`C:\Users\wangq\Documents\Codex\knowledge-vault\work\internal\强哥的思想宇宙\GPT成果\2026-06-16-核心模型抽取样板 v0.1.md`
|
||
|
||
本仓库本地化规则位于：
|
||
|
||
- `docs/GPT_PLAN_LOCALIZATION_PROTOCOL.md`
|
||
- `docs/MODEL_EXTRACTION_RULES.md`
|
||
- `docs/MODEL_EXTRACTION_WORKFLOW.md`
|
||
|
||
## 核心判断
|
||
|
||
GPT 规划不是本地可直接执行规则。
|
||
|
||
Codex 必须先把 GPT 规划转成：
|
||
|
||
- local rules
|
||
- schemas
|
||
- workflow gates
|
||
- indexes
|
||
- validation tools
|
||
- owner review points
|
||
|
||
再进入内容抽取。
|
||
|
||
## 模型抽取链路
|
||
|
||
```text
|
||
原始文章 / 代表文本
|
||
-> source article record
|
||
-> source evidence excerpts
|
||
-> human-readable Markdown card
|
||
-> machine-readable JSON model card
|
||
-> regression cases
|
||
-> selector examples or routing rules
|
||
-> model/card indexes
|
||
-> validation report
|
||
```
|
||
|
||
## 双工模式
|
||
|
||
模型抽取必须分为两条线：
|
||
|
||
1. workflow/tooling track
|
||
2. content extraction track
|
||
|
||
内容抽取不能吸收缺失的规则、schema、validator、index 或 workflow 工作。
|
||
|
||
## 当前本地样板
|
||
|
||
当前本地样板模型：
|
||
|
||
- QPI
|
||
- 思想考古学
|
||
|
||
当前验证命令：
|
||
|
||
```powershell
|
||
python scripts\rebuild_indexes.py --check
|
||
python -m unittest discover -s tests -p "test*.py" -v
|
||
python scripts\check_card_contract.py
|
||
python scripts\validate_model_library.py
|
||
python scripts\run_selector_demo.py
|
||
python scripts\run_selector_regression.py
|
||
python scripts\check_model_card_sync.py
|
||
```
|
||
|
||
## v0.2 内容稳定化补充
|
||
|
||
工程 contract 通过不等于内容稳定。
|
||
|
||
字段级 evidence coverage 必须区分：
|
||
|
||
- `direct_source`
|
||
- `derived_from_source`
|
||
- `product_decision`
|
||
- `red_team_inference`
|
||
- `owner_decision`
|
||
|
||
source excerpt 必须标记：
|
||
|
||
- `quote_status`
|
||
- `source_location`
|
||
|
||
`quote_status=exact` 表示准确摘录；压缩摘录必须使用 `quote_status=condensed` 并在 notes 中说明。
|
||
|
||
回归测试在内容稳定化阶段每个核心模型至少 15 条，并覆盖：
|
||
|
||
- positive
|
||
- boundary
|
||
- misuse
|
||
- no_call
|
||
- selector_gate
|
||
- pipeline
|
||
|
||
selector v0.2 仍然是规则 selector，不引入 LLM。它必须输出 selected/rejected models、score breakdown、penalties 和 no-call 状态。
|