Generates synthetic datasets for chained arithmetic completion. The agent must compute multiple sequentially dependent arithmetic expressions, where each line's result becomes a substrate (or operand) for subsequent lines, requiring faithful arithmetic execution and write-only completion.
Each sample pairs a task (first frame + prompt describing what needs to happen) with its ground truth solution (final frame showing the result + video demonstrating how to achieve it). This structure enables both model evaluation and training.
| Property | Value |
|---|---|
| Task ID | Multi-21 |
| Task | Chained Math Calculation |
| Category | Algorithmic Execution |
| Resolution | 1024×1024 px |
| FPS | 16 fps |
| Duration | varies |
| Output | PNG images + MP4 video |
# 1. Clone the repository
git clone https://github.com/VBVR-DataFactory/Multi-21_chained_math_calculation_data-generator.git
cd Multi-21_chained_math_calculation_data-generator
# 2. Create and activate virtual environment
python3 -m venv venv
source venv/bin/activate # On Windows: venv\Scripts\activate
# 3. Install dependencies
pip install --upgrade pip
pip install -r requirements.txt
pip install -e .# Generate 50 samples
python examples/generate.py --num-samples 50
# Reproducible generation with seed
python examples/generate.py --num-samples 50 --seed 42
# Custom output directory
python examples/generate.py --num-samples 100 --output data/my_dataset
# Without videos (faster, images only)
python examples/generate.py --num-samples 50 --no-videos| Argument | Description |
|---|---|
--num-samples |
Number of tasks to generate (required) |
--output |
Output directory (default: data/questions) |
--seed |
Random seed for reproducibility |
--no-videos |
Skip video generation (images only) |
[Scenario] The image displays a sequence of stacked addition problems from top to bottom. The first problem has two operands, and subsequent problems have one missing operand.
[Rules]
1. The equations must be solved sequentially from top to bottom.
2. The left operand of each row (after the first) is exactly the calculated answer from the previous row.
[Task] Generate a video animating the step-by-step calculation of these chained equations from top to bottom. Fill in the correct intermediate result and the final answer after each '=' sign while leaving all other text unchanged.
 |
 |
 |
| Initial Frame Math expressions with empty results |
Animation Each line filled with computed result |
Final Frame All expressions completed |
Compute each arithmetic expression in the displayed text and fill in the result after the equals sign, without modifying any other characters in the text.
- Expressions: Multiple arithmetic lines (e.g.,
123 + 456 = ?). - Operators: Mix of
+,-,×,÷, possibly with parentheses. - Number range: Multi-digit operands (chosen for non-trivial mental computation).
- Strict typography: Only the answer slot is mutated; surrounding text, formatting, and unrelated lines must remain pixel-stable.
- Video reveal: Each expression's answer is filled in sequentially, providing a per-line correctness signal.
- Faithful arithmetic execution: Tests pure computational fidelity (no semantic ambiguity, no plan).
- Write-stability constraint: The model must update only the answer slots, treating the rest of the frame as immutable — a strong test of selective generation.
- Sequential dependency option: When chained, each line's result feeds the next, so an early arithmetic mistake propagates and is observable.
- Frame-by-frame evaluation: The video isolates each fill-in step, supporting fine-grained scoring.
data/questions/Multi-21_chained_math_calculation_data-generator_task/Multi-21_chained_math_calculation_data-generator_00000000/
├── first_frame.png # Expressions with blank '=' slots
├── final_frame.png # All slots filled with correct results
├── prompt.txt # Task instruction
├── ground_truth.mp4 # Animation of stepwise filling
└── question_metadata.json # Standardized VBVR task metadata
File specifications:
- Images: 1024×1024 PNG format
- Video: MP4 format, 16 fps, H.264 + yuv420p
- Metadata: VBVR canonical schema with
task_id,vbvr_task_code,media,parameters
arithmetic math-calculation chained-computation algorithmic-execution write-stability multi-step-reasoning
Part of the 36-Task Long-Horizon Multi-Step Reasoning Benchmark.