Skip to content

agent_sop.md

Latest commit

 

History

History
55 lines (45 loc) · 2.83 KB

agent_sop.md

File metadata and controls

55 lines (45 loc) · 2.83 KB
Raw

PredActor agent and contributor guide

This is the compact navigation layer for coding agents. The surrounding guides remain the human-readable instructions for collection, dyeing, training, evaluation, and deployment. Keep this file synchronized from PegasusMoDye/doc/predactor_agent_sop.md when the shared workflow changes.

Choose the requested path

Request Read first Entry point
Quick evaluation README.md, docs/evaluation.md uv run --locked predactor-eval
Data collection README.md, collection and dyeing guides PredActor/collection/collect_dataset.sh
Policy training README.md, docs/training.md PredActor/train.py
Trained-model evaluation docs/training.md, docs/evaluation.md predactor-eval --checkpoint ...
Deployment docs/deployment.md, PredActor/diffusion_deploy/README.md export, build, and choose a launch route

Before running a command, resolve the environment, artifact, config, and checkpoint from the linked guide. Do not guess from a filename or silently substitute another experiment. If an input is missing, report it clearly.

Source-first workflow

  1. Work on a task branch and preserve unrelated local changes.
  2. Read the applicable human guide and this guide before editing or launching.
  3. Treat the PegasusMoDye repository as the source of truth for shared docs and mappings; this repository is the public export.
  4. Verify versions, paths, and artifact identities before an expensive command. Prefer --help, a dry-run, or a smoke test first.
  5. Keep logs, checkpoints, and large datasets in their documented artifact locations, outside the Git source tree when the guide says so.
  6. Report the command, working directory, config/checkpoint paths, environment assumptions, validation level, output path, and any blocker.

The learned weights are maintained separately in MasterYip/PredActor_Artifacts. Do not claim reproducibility when a required dataset or weight is unavailable.

Contributing

Star the project when it is useful, open an issue for a reproducible bug or missing input, fork for independent work, and submit a focused pull request. These are ordinary collaboration routes; publishing to GitHub or Hugging Face still requires the project owner's approval.