RARE-Negative

RARE-Negative contains the final sample-level Negative checkpoints from RARE: Reinforcement Learning with Adaptive Rubric Evolution for Open-Ended Generation. The model generates complete Chinese video scripts from short creative briefs.

Publication metadata and the final BibTeX entry will be added when they are available.

Model description

RARE adapts the rubric used as an RL reward to the current policy. Each query maintains its own rubric. When one criterion gives full credit to all eight on-policy scripts, RARE replaces only that saturated criterion and immediately rescores the same group. The Negative variant writes replacement criteria around observable flaws in the current outputs, with more severe flaws receiving lower scores. Unchanged criteria retain their scores and structure.

This repository packages three independent full checkpoints, one for each production style. They are not adapters and are not intended to be combined.

Subfolder	Production style	Final RL step	Dataset configuration
2d_animation	2D animation	132	2d_animation
3d_stopmotion	3D and stop-motion animation	108	3d_stopmotion
live_action	Live action	94	live_action

Results

Scores below use the paper's held-out script-generation evaluation. Each style is averaged over its test queries, and Avg. equally averages the three styles. The evaluation protocol and judge details are described in the paper.

Model	2D animation	3D/stop-motion	Live action	Avg.
Base	65.51	64.63	60.83	63.66
Sample-Static	74.70	72.94	68.29	71.98
RARE (Sample-level, Negative)	78.25	76.84	72.31	75.80

Negative improves over the matched Sample-Static baseline by 3.82 points on average. Across 3,336 valid criterion switches, 80.0% reopen score headroom and 78.2% restore within-group score variance.

Training details

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Optimization: GRPO, two epochs per production style
• Global batch size: 32
• Rollouts per query: 8
• Learning rate: 1e-6
• Maximum prompt length: 4,096 tokens
• Maximum response length: 8,192 tokens
• Training seed: 1

All three styles use the same optimization configuration and differ only in their training data and resulting final step.

Reproduction and checkpoint use

The exact training launchers, optimization settings, task-specific schedules, and validation behavior are documented in the RARE repository. The corresponding configuration in EurekaTian/RARE preserves the released system instruction and query format. Each subfolder is an inference-only full checkpoint; optimizer and trainer state are not included. This card does not add a separate convenience inference command that was not part of the released pipeline.

Intended use and limitations

The checkpoints are intended for research on open-ended generation, rubric-based rewards, and long-form Chinese screenplay generation. They were trained separately for the three listed production styles; select the closest style instead of treating the subfolders as an ensemble.

The models may produce inaccurate, repetitive, biased, or unsafe content. The reported scores rely on the paper's model-based evaluation protocol and should not be interpreted as guarantees for production use or other domains.

License

The model weights are released under the Apache License 2.0, following the base model license. The RARE query-rubric dataset is released separately under CC BY 4.0 and remains subject to its dataset card and source-data notes.

Citation

Please cite the paper through its OpenReview record. The canonical BibTeX entry will be added after the publication metadata is finalized.