PRISM: Bilingual Benchmark for Programmatic Teaching-Video Generation with Manim

10,372 examples form a benchmark for generating programmatic teaching videos using the Manim library. The dataset contains 5,199 English and 5,173 Chinese examples, each pairing a natural-language instruction with a reference Manim answer. It was created by 'posprivacy' and last updated on 2026-05-03.

Use Cases

• Training models to generate Manim code based on natural-language instructions.
• Benchmarking the performance of text-to-animation or code-generation models.
• Studying cross-lingual instruction-following capabilities in English and Chinese.

Strengths

• Contains 10,372 total examples, providing a substantial benchmark size.
• Includes a balanced bilingual split with 5,199 English and 5,173 Chinese examples.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the main data file is unknown, which may limit suitability assessment.

Provenance

Source

huggingface

Collection Method

Likely created as a benchmark for programmatic video generation.

Freshness

Last updated 2026-05-03 04:11:20; freshness should be verified.