--- license: apache-2.0 language: - zh library_name: transformers pipeline_tag: text-generation base_model: Qwen/Qwen3-4B-Instruct-2507 datasets: - EurekaTian/RARE tags: - reinforcement-learning - grpo - adaptive-rubric - script-generation --- # RARE-Task-Level This repository contains the final **task-level adaptive rubric** checkpoints from **RARE: Reinforcement Learning with Adaptive Rubric Evolution for Open-Ended Generation**. Within each production style, all queries share a six-criterion rubric that is periodically revised from cross-query on-policy evidence while preserving its IDs, order, score maxima, and reward aggregation.

| Subfolder | Production style | Final RL step | |---|---|---:| | `2d_animation` | 2D animation | 132 | | `3d_stopmotion` | 3D and stop-motion animation | 108 | | `live_action` | Live action | 94 | The held-out scores are 76.27, 75.56, and 69.84 for the three styles, respectively, for an equally weighted average of **73.89**. This is a 1.08-point gain over the matched Task-Static baseline. Each subfolder is an independent inference-only full checkpoint derived from `Qwen/Qwen3-4B-Instruct-2507`; optimizer and trainer state are not included. The exact launcher and settings are documented in the [RARE repository](https://github.com/Eureka-Maggie/RARE#training). Use the corresponding configuration from [EurekaTian/RARE](https://huggingface.co/datasets/EurekaTian/RARE) for the released system instruction and query format. This card does not add a separate convenience inference command that was not part of the released pipeline. The weights are licensed under Apache-2.0; the dataset is separately licensed under CC BY 4.0. The canonical paper citation will be added after publication metadata is finalized.