observation也可最先从调用普通deepseek的api开始，然后再是上服务器用小模型去观察，发现确实有这个现象，然后我们就lora finetu…
- Stable ID:
codex-user-prompt:9c081859b7858b00
- Source kind:
codex-session-user
- Category:
research-workflow
- Timestamp:
2026-05-11T10:41:02.585Z
- Semantic hash:
9c081859b7858b00e8b7de6f408b0d8aba12c1a2a494f62e9dc25ffd549893da
- Public handling: selected full-text prompt with secret filtering.
Prompt Text
observation也可最先从调用普通deepseek的api开始，然后再是上服务器用小模型去观察，发现确实有这个现象，然后我们就lora finetune（我不知道相关术语，反正你看别的论文是怎么弄的）这个模型，然后也要在别人的baseline上看到他仍然有我们发现的这个痛点，然后后面通过训练对比其他baseline，证实确实比别人的要好
Reuse Notes
- EXTRACTED: This is a selected Codex prompt or automation prompt from the local Codex corpus.
- INFERRED: Future agents can reuse its structure, constraints, and acceptance criteria when creating similar Codex workflows.