---
pretty_name: Physical-ICL WISA5000 TI2V Data v1
---

# Physical-ICL WISA5000 TI2V Data v1

This directory is the metadata-only train/test package derived from the frozen Physics Card v7 / alpha-3 WISA5000 retrieval pool. It is intended for text-and-image-to-video Physical-ICL experiments in which a model is trained on directed physical demonstration pairs and evaluated on 100 held-out query videos with one reviewed oracle demo per query.

No video or first-frame files are included. Media are resolved by `sample_id`; users must obtain the corresponding WISA videos under their applicable access and licensing terms. The package does not grant redistribution rights for the source media.

## Counts

| Artifact | Count |
| --- | ---: |
| Required unique videos | 4,118 |
| Metadata rows | 4,118 |
| Train pairs | 37,415 |
| Held-out test queries | 100 |
| Test query-demo pairs | 100 |
| Unique test demos | 99 |

## Files

- `video_metadata.jsonl`: one row per required video with the original WISA caption, the full successful Physics Card, and train/test roles.
- `train_pairs.jsonl`: directed training relations with retrieval scores.
- `test100_1demo.jsonl`: one reviewed oracle demonstration per held-out test query.
- `test100_ids_final.txt`: the 100 held-out query IDs.
- `required_wisa_ids.txt`: exact union of every query/demo endpoint referenced by train or test.
- `split_summary_final.json`: split construction and declared invariants.
- `dataset_summary.json`: storage, text, Physics Card, and count policies.
- `retrieval_summary_clean_v21.json`: frozen retrieval configuration and aggregate statistics.
- `VALIDATION_REPORT.json`: independent structural and leakage audit performed before upload.
- `SHA256SUMS`: SHA-256 for every file except the checksum manifest itself.

## Split semantics

The 100 test query IDs are removed as both queries and demos from training. Starting from 39,260 source relations, removing every edge incident to a test query yields exactly 37,415 train pairs; 1,845 relations are removed.

At evaluation time, each test query receives its listed `demo_id`. Every test demo belongs to the train bank and is not itself a test target. This is an **oracle Physical-ICL demo evaluation** rather than an evaluation of an automatic demo selector.

Do not reconstruct a random pair-row split. Use the supplied split directly, and never use a test query video as training input or context.

## TI2V media preparation

Resolve every ID in `required_wisa_ids.txt` to its source video. The package intentionally does not prescribe or include first-frame extraction. Record the chosen decoding backend, frame index/timestamp, color conversion, resizing, and cropping policy so train and test inputs are reproducible.

## Known limitation: visual copy risk

Video-ID leakage is eliminated, and text screening found no exact query-demo caption, event-summary, or raw-event-description duplicates. However, all 100 test pairs remain marked `visual_copy_risk_status = unreviewed`. Text screening is not a substitute for perceptual/video near-duplicate detection. Do not claim near-duplicate-safe generalization until the source videos have been checked visually or clustered with a suitable video fingerprint/embedding method.

The annotation and retrieval workflow is available in the [`physical_icl/`](https://github.com/chengye629/Physical-VICL/tree/main/physical_icl) module.