GraphNet 自动样本抽取 Agent 设计 #602
Dayuxiaoshui
started this conversation in
Ideas
Replies: 4 comments 2 replies
-
|
run_model.py也应该需要入库才对的。收集足够多的run_model.py,能够帮助自动抓取agent不断改进。 |
Beta Was this translation helpful? Give feedback.
1 reply
-
|
量化指标:成功生成并验证通过的样本数 / 总尝试数 |
Beta Was this translation helpful? Give feedback.
0 replies
-
|
还得有一个查重的环节。防止和已有的模型重复了。 |
Beta Was this translation helpful? Give feedback.
0 replies
-
|
ADT形式化描述。
从上面的定义可以看出来,当前的类型定义不够明确。 |
Beta Was this translation helpful? Give feedback.
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
GraphNet 自动样本抽取 Agent 设计方案
1. 问题描述
run_model.py,效率低且难以应对海量异构模型。ModelID → Sample2. 方案设计
Agent 采用模块化设计,通过状态流转实现自动化。
2.1 数据结构
run_model.py脚本2.2 核心操作
ModelID → Model- 模型下载与同步Model → Metadata- 解析 config.json 提取元数据Metadata → Script- 混合策略:优先 LLM 生成,失败则使用通用模板Script → Sample- 在子进程中执行脚本并提取计算图Sample → bool- 基于 graph_hash 检查是否与已有样本重复Sample → bool- 样本合法性校验Script → KnowledgeBase- 将生成的 run_model.py 保存到样本目录,形成知识库2.3 架构流程图
graph LR ID[HF Model ID] -->|Fetch| M[Local Model] M -->|Analyze| Meta[Metadata] subgraph Codegen [CodeGen Strategy] direction TB Meta --> LLM[LLM Engine] Meta --> Tpl[Template] end LLM --> S[run_model.py] Tpl --> S S -->|Extract| G[Sample] G -->|Deduplicate| Check{Duplicate?} Check -->|Yes| Skip[Skip Existing] Check -->|No| Verify[Verify] Verify --> Result{Success?} Result -->|Yes| Archive[Archive Script] Archive --> KB[KnowledgeBase] KB -.Feedback.-> CodegenBeta Was this translation helpful? Give feedback.
All reactions