StepCountQA-RL-Dense-Plus Dataset

Dataset Description

StepCountQA-RL-Dense-Plus is a carefully filtered subset of StepCountQA-RL, containing complete reasoning chains where the final count is between 11 and 50.

Key Feature: Each sequence includes ALL reasoning steps from count=1 to the final count (11-50), making it ideal for training models on dense counting scenarios with complete reasoning processes.

Dataset Statistics

• Training Samples: 192,980
• Sequences: ~7,800 complete reasoning chains
• Count Range: 11-50 (final count)
• Average Steps per Sequence: ~24 steps

Data Structure

Complete Reasoning Chain Format

Each counting task contains a full reasoning chain from the first to the last point:

image.jpg          -> count=1, {"point_2d": [x1, y1], "label": "object", "count_number": 1}
image_1.jpg        -> count=2, {"point_2d": [x2, y2], "label": "object", "count_number": 2}
image_2.jpg        -> count=3, {"point_2d": [x3, y3], "label": "object", "count_number": 3}
...
image_N.jpg        -> count=N+1 (where N+1 is between 11-50)

Data Fields

• images: A sequence of images (typically one image per sample)
• problem: Question text with reasoning instructions (<image>\nHow many [objects] are in the image?\n...)
• answer:
  • During reasoning steps: JSON format {"point_2d": [x, y], "label": "...", "count_number": N}
  • Final answer: Simple number string "N"

Dataset Characteristics

1. Complete Reasoning Chains

• Every sequence starts from count=1
• Includes all intermediate steps
• Ends with final count between 11-50

2. Dense Counting Scenarios

• Focus on moderately dense object counts (11-50 objects)
• Suitable for training on challenging counting tasks
• Balances complexity and tractability

3. Diverse Object Types

• People, vehicles, everyday objects
• Fine-grained object parts (hands, heads, etc.)
• Various scenes and contexts

Usage Example

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("MING-ZCH/StepCountQA-RL-Dense-Plus")

# Access training data
train_data = dataset["train"]

# View a sample
sample = train_data[0]
print(sample['problem'])
print(sample['answer'])
# The answer may be JSON (intermediate step) or a number (final answer)

Training Recommendations

This dataset is particularly useful for:

• Incremental counting models: Learn to count step-by-step
• Dense object detection: Train on moderately crowded scenes
• Reasoning consistency: Ensure models maintain coherent reasoning chains
• Point-based annotation: Learn precise spatial localization

Citation

If you use this dataset, please cite the original StepCountQA-RL dataset.

License

Follows the same license as the original StepCountQA-RL dataset.