RARE Query-Rubric Dataset

This dataset contains the query-specific rubric annotations used by RARE for reinforcement learning on complete screenplay generation. It provides 5,989 query-rubric pairs grouped by 930 source videos and covers three production styles: 2D animation, 3D/stop-motion, and live action.

[Code] [Paper]

Dataset summary

Configuration	Train	Test	Source-video groups
2d_animation	2,125	237	370
3d_stopmotion	1,759	192	302
live_action	1,509	167	258
Total	5,393	596	930

All queries derived from the same source video remain on the same side of the train-test boundary. Source-video titles and uploader handles have been replaced by stable, non-identifying IDs. The original videos are not included in this release.

Loading

Load an individual production style with datasets:

from datasets import load_dataset

dataset = load_dataset("EurekaTian/RARE", "2d_animation")
print(dataset["train"][0])

The repository also preserves the exact Parquet layout expected by the released training code:

data/scripts/video_corpus_manual_v1_1b_20260807/
├── 2d_animation/{train,test}.parquet
├── 3d_stopmotion/{train,test}.parquet
└── live_action/{train,test}.parquet

Schema

Field	Description
problem	The screenplay-generation query.
answer	Reserved compatibility field; empty in this annotation release.
difficulty	Compatibility metadata from the training format.
query_id	Stable integer query identifier.
criteria	Serialized query-specific rubric.
data_source	Production-style data-source identifier.
prompt	Training-ready system and user chat messages.
ability	Task label (script).
reward_model	Compatibility metadata for the reward pipeline.
extra_info	Query, rubric, query index, anonymized video-group ID, and query UID.

Rubrics contain three to eight criteria with explicit checks, scoring rules, and maximum scores. They initialize the sample-level static and dynamic RARE training variants. Task-level initial rubrics are distributed with the code.

Data construction

The source collection was deduplicated into 930 video groups. Each video was used to derive multiple screenplay-generation queries and query-specific rubrics, followed by a group-disjoint 90/10 train-test split within each production style. The released rows were checked for absolute paths, credentials, email addresses, internal service names, and source-identifying video metadata.

The exact file counts and SHA-256 digests are recorded in release_manifest.json.

Intended use and limitations

The dataset is intended for research on rubric-based reward modeling, open-ended generation, and reinforcement learning. It is primarily Chinese and focuses on screenplay generation in three production styles. Queries and rubrics were produced with model assistance and may contain cultural, stylistic, or evaluation biases. Users should inspect the criteria before adapting them to other domains or languages.

The release contains annotations only. It does not grant rights to any source video, model checkpoint, or third-party training framework.

License

The released query-rubric annotations are available under CC BY 4.0. The RARE code uses the Apache License 2.0 and is distributed separately.

Citation

The paper citation will be added when the public paper record is available. For now, please cite the dataset repository and the associated code release.