Attack-WAM Watermark Artifacts

This repository stores generated watermark artifacts for RoboTwin and four LIBERO suites. It separates full-data runs from dataset-ablation runs and uses one stable layout across machines.

Coverage

• Benchmarks: libero, robotwin
• LIBERO suites: long, goal, object, spatial
• RoboTwin suite key: main
• Regimes: full, ablation/episode_pool, ablation/task_fraction
• Variants:
  • direct_A1V0
  • direct_A0V1
  • direct_A1V1
  • de_proj_A1V0
  • de_proj_A0V1
  • bisect_A1V1

Canonical paths

artifacts/full/libero/<long|goal|object|spatial>/<variant>/
artifacts/full/robotwin/main/<variant>/

artifacts/ablation/episode_pool/libero/<suite>/top_<10|20|30|40|50>/<variant>/
artifacts/ablation/episode_pool/robotwin/main/top_<10|20|30|40|50>/<variant>/

artifacts/ablation/task_fraction/libero/<suite>/p<10..90>/<variant>/
artifacts/ablation/task_fraction/robotwin/main/p<10..90>/<variant>/

Each leaf directory follows:

<variant>/
├── metadata.json             # required
├── top1.npy                  # required
├── top2.npy                  # recommended
├── top3.npy                  # recommended
├── top4.npy                  # recommended
├── top5.npy                  # recommended
├── final.npy                 # recommended
└── optimization.jsonl        # optional

top1.npy is the default artifact for evaluation. final.npy is the last optimization state and is not necessarily the best state.

See SPEC.md for naming, metadata, and upload requirements.

Validation

Run before every upload:

python tools/validate_repository.py /path/to/local/repository

Use --require-complete-full to require all 30 full-data combinations (five dataset suites times six variants).

Repository manifest

manifest.json is the machine-readable index. Uploading machines must update it after placing artifacts. Paths in the manifest are repository-relative and must never contain local filesystem paths or credentials.