Create reconstruction.md

metadata/reconstruction.md

@@ -0,0 +1,29 @@
+# SFT-Eval Split Reconstruction
+
+## Relevant Assets
+
+- QA rows for `eval split` and `sft split` in splits/
+- Split manifests and split counts in splits/
+
+## How to inteprete the QA rows
+
+Each row exposes three reconstruction-oriented fields:
+
+- `visual_ref`
+- `source_episode_ref`
+- `reconstruction_key_json`
+
+These are designed to be stable, path-free references. They are sufficient to map a released item back to the corresponding source episode or recording once you have obtained the upstream dataset under its original terms.
+
+## Source-specific units
+
+- `GM-100`: grouped by `(task_id, episode_id)`.
+- `RH20T`: grouped by `recording_id`.
+- `REASSEMBLE`: grouped by `recording_id`.
+- `AIST-Bimanual`: grouped by `recording_id`.
+
+## Expected workflow
+
+1. Obtain upstream datasets under their original terms.
+2. Use `source`, `source_task_id`, `source_unit_id`, `camera`, and `frame_indices_json` from the public release.
+3. Reconstruct the required visual inputs with the source-specific evaluation scripts in the main codebase.