---
task_categories:
  - visual-question-answering
tags:
  - chronoreplay
  - spatial-reasoning
  - causal-video
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.parquet
---

# ChronoReplay-Train

Portable training annotations for **ChronoReplay: One Answer Interface Is Not
Enough for Streaming Spatial Intelligence**.

This release contains all 39,561 selected serialized training records with
their original questions, answers, order, loss weights, and supervised-turn
masks. It contains **annotations and relative media references, not the
referenced image/video files**.

| Supervision | Serialized records |
|---|---:|
| Categorical preservation | 30,635 |
| Numeric geometry / camera motion | 2,545 |
| OST causal episodes | 6,381 |
| Total | 39,561 |

The OST episodes contain 45,006 supervised trajectory turns. The full pool
references 946,256 unique media paths. This is a reference count; it does not
assert that all media are present in this repository or independently mirrored.

## Files

- `data/train.parquet`: portable native training records.
- `manifests/replay_selection.json`: original IDs, ordering, text hashes,
  observation counts, weights, and supervision masks.
- `manifests/media_paths.jsonl.gz`: unique relative paths referenced by the data.
- `release_manifest.json`: conversion checks, component counts, and data hashes.

Server-specific media prefixes were replaced with relative paths under `media/`.
Internal conversion-certificate columns were omitted. Every native question
and target was checked against the frozen per-record text hash; identifiers,
ordering, weights, masks, and image counts were also checked for all records.
The portable Parquet has a new file hash because its paths and schema differ
from the original server artifact. No new samples or target answers were generated.

## Use with the code release

Download the annotations and obtain the source media under their applicable
terms. Arrange media under the relative paths listed in the records. From the
ChronoReplay code release:

```bash
python -m scripts.materialize_frozen \
  --candidate downloaded/data/train.parquet \
  --selection downloaded/manifests/replay_selection.json \
  --media-root downloaded \
  --output data/replay
```

The materializer verifies the frozen records and decodes every referenced
image before declaring the pool ready. This annotations-only upload is not
a self-contained media distribution.

## Sources and terms

Sources include the [OST-Bench training release](https://huggingface.co/datasets/rbler/OST-Bench),
source-annotated categorical preservation examples, and
[ARKitScenes](https://github.com/apple/ARKitScenes) camera-rotation and
corrected-camera geometry supervision. Original source terms continue to apply.
The OST training build records CC-BY-NC-4.0 for that component; this statement
does not assign that license to the entire mixed-source pool.

The [ChronoReplay collection](https://huggingface.co/collections/mis2/chronoreplay-6adc88dc1e468c2d2ae9bb44)
links these annotations and the corresponding model; the preparation scripts
are in the [code repository](https://github.com/yihaohu0118/ChronoReplay).
This card does not assign a new blanket license to third-party source material.
Media availability and current pixel identity have not been revalidated by
this annotation conversion. No evaluation predictions or score artifacts are bundled.

See `CITATION.cff` for citation information.