Cereloop/manim_animation_dataset

Primary Maths Manim Instruction Dataset

This is a synthetic and model-repair instruction-tuning dataset for training coding models to generate Manim Community Edition Python code for primary-school mathematics animation videos.

The current final dataset contains 6,023 chat-formatted examples covering:

• Fractions
• Place Value
• Multiplication and Division

It is designed for fine-tuning models such as Qwen2.5-Coder-Instruct to produce:

• executable Manim animation code
• simple maths storyboards
• Manim code from storyboards
• corrected Manim code from broken or weak examples
• fixed code from Manim/Python error messages
• improved outputs that better follow layout, duration, font, visual-model, exact-count, and render-error instructions

Dataset Files

Recommended file for Unsloth/ShareGPT training:

primary_maths_manim_sharegpt_final_6023.jsonl

Qwen-style messages version:

primary_maths_manim_qwen_messages_final_6023.jsonl

The latest 800-example targeted repair add-on is also provided separately:

manim_error_repair_dataset_0800.jsonl
manim_error_repair_qwen_messages_0800.jsonl
manim_error_repair_sharegpt_0800.jsonl

Validation summary for the latest add-on:

manim_error_repair_validation_summary.json

Dataset Format

The Qwen version uses messages format:

{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}

The ShareGPT/Unsloth version uses conversations format:

{
  "conversations": [
    {"from": "system", "value": "..."},
    {"from": "human", "value": "..."},
    {"from": "gpt", "value": "..."}
  ]
}

The assistant responses contain either complete Manim Python code, a numbered storyboard, or corrected Manim code, depending on the task.

Dataset Size

Total examples: 6,023

Dataset composition:

• Foundation topic dataset: 2,700 examples
• Instruction-following repair dataset: 600 examples
• Instruction-gap repair dataset: 1,000 examples
• Real generation repair dataset from logged model failures: 123 examples
• Outcome-boost dataset: 800 examples
• Targeted Manim error repair dataset: 800 examples

Overall topic distribution:

• Multiplication and Division: 2,147
• Fractions: 2,003
• Place Value: 1,873

Overall grade distribution:

• Primary 2: 973
• Primary 3: 1,261
• Primary 4: 1,182
• Primary 5: 1,322
• Primary 6: 1,285

Foundation topic distribution:

• Fractions: 900
• Place Value: 900
• Multiplication and Division: 900

Foundation task distribution:

• topic_to_manim_code: 1,080
• topic_to_storyboard: 405
• storyboard_to_manim_code: 540
• bad_code_to_fixed_code: 405
• error_message_to_fixed_code: 270

The 800-example targeted Manim error repair add-on focuses on observed model failures:

• hallucinated Manim APIs such as DropInEffect, DrawChars, and arrange_clockwise
• invalid object methods such as Sector.add_to_edge and pairs.rect
• SurroundingRectangle keyword errors
• undefined variables such as i and answer_label
• fragile LaTeX that fails during rendering
• wrong multiplication array counts
• wrong division grouping
• decimal place-value label mistakes
• duration control when animations collapse to about one second
• font/readability repairs that should preserve the original scene structure

Targeted repair add-on task distribution:

• error_message_to_fixed_code: 370
• bad_code_to_fixed_code: 370
• topic_to_manim_code: 40
• storyboard_to_manim_code: 20

Targeted repair add-on skill distribution:

• replace_hallucinated_manim_api: 160
• repair_wrong_visual_math: 160
• respect_requested_duration: 120
• fix_runtime_error: 110
• fix_keyword_or_type_error: 100
• minimal_font_readability_edit: 90
• safe_manim_api_generation: 40
• storyboard_to_safe_code: 20

Validation

The dataset was checked for:

• valid JSONL formatting
• correct chat message structure
• unique chat prompts and outputs in generated repair blocks
• from manim import * in code outputs
• exactly one Scene class in code outputs
• no external assets
• Python syntax validity
• explicit timing patterns in duration-repair examples
• no fake Manim API tokens in corrected targeted-repair outputs

The original 2,295 foundation Manim code outputs were rendered successfully with Manim Community Edition using Docker.

The latest 800-example targeted Manim error repair add-on passed static validation:

• 800 valid rows
• 800 unique IDs
• 0 duplicate chat prompts
• 0 duplicate outputs
• all corrected code outputs parse as Python
• all corrected code outputs define exactly one Scene
• all corrected code outputs start with from manim import *
• all corrected code outputs include explicit run_time pacing
• no external assets
• no known fake API tokens in corrected outputs

A stratified Docker render sample of the latest repair add-on passed:

24 tested
24 passed
0 failed
Docker image: manimcommunity/manim:stable

The full merged dataset was assembled from validated source blocks. Some animations may still benefit from human visual review for polish and pedagogy, even when the code renders successfully.

Intended Use

This dataset is intended for supervised fine-tuning of coding or instruction-following models to improve performance on primary-school maths animation generation using Manim.

It is especially suited for models that need to generate visual educational content from natural-language teacher prompts and revise flawed Manim outputs.

Recommended training use cases:

• fine-tuning Qwen2.5-Coder-Instruct-style models
• training Manim code generation behavior
• improving instruction following for visual constraints
• teaching model repair from Manim render errors
• improving exact object counts for arrays, groups, fractions, and place-value charts
• reducing hallucinated Manim APIs
• collecting future preference data from generated outputs

Limitations

This dataset is mostly synthetic, with a smaller set of real model-generation repair examples. Although the code was validated for syntax and sampled for rendering, some animations may still require human review for visual polish, pedagogy, or exact classroom suitability.

The dataset is focused on primary-school fractions, place value, multiplication, and division. It should not be assumed to generalize strongly to unrelated subjects without additional training data.