---
dataset_info:
  features:
  - name: images
    sequence: image
  - name: problem
    dtype: string
  - name: answer
    dtype: string
  splits:
  - name: train
    num_examples: 39230
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# StepCountQA-RL-Traj_11_50_Combined

## Dataset Description

StepCountQA-RL-Traj_11_50_Combined is a unified visual counting RL training dataset covering object counts from **11 to 50**. It merges two complementary 11-50 datasets into a single collection:

| Source | Samples | Description |
|--------|---------|-------------|
| StepCountQA-RL-Traj_11_50 | 9,508 | `<point>` JSON annotation trajectories, converted to numeric |
| StepCountQA-RL-Traj_11_50_NumericOnly | 29,722 | Pure numeric answer from SFT trajectories |
| **Combined** | **39,230** | Merged, deduplicated (0 overlap) |

All samples pair an original (non-annotated) image with a counting question and a **pure numeric answer string**.

## Dataset Structure

### Data Fields

| Field | Type | Description |
|-------|------|-------------|
| `images` | `Sequence(Image)` | One original image per sample |
| `problem` | `string` | Counting question with `<image>` tag |
| `answer` | `string` | Numeric count as string (e.g., `"27"`) |

### Format Example

```python
{
  "images": [<PIL.Image>],
  "problem": "<image>\nHow many people are in the image?",
  "answer": "39"
}
```

## Dataset Statistics

- **Total samples**: 39,230
- **Count range**: 11 – 50
- **Unique images**: 39,230 (zero duplicates verified)
- **Total size**: ~8.95 GB (4 parquet shards)

### Count Distribution

| Count | Samples | Count | Samples |
|-------|---------|-------|---------|
| 11 | 2,537 | 31 | 660 |
| 12 | 3,297 | 32 | 676 |
| 13 | 1,863 | 33 | 591 |
| 14 | 2,147 | 34 | 558 |
| 15 | 2,145 | 35 | 621 |
| 16 | 2,011 | 36 | 687 |
| 17 | 1,372 | 37 | 469 |
| 18 | 1,732 | 38 | 484 |
| 19 | 1,235 | 39 | 464 |
| 20 | 1,621 | 40 | 670 |
| 21 | 1,136 | 41 | 425 |
| 22 | 1,072 | 42 | 474 |
| 23 | 967 | 43 | 359 |
| 24 | 1,182 | 44 | 383 |
| 25 | 976 | 45 | 424 |
| 26 | 907 | 46 | 364 |
| 27 | 753 | 47 | 323 |
| 28 | 879 | 48 | 432 |
| 29 | 627 | 49 | 341 |
| 30 | 940 | 50 | 426 |

## Comparison with Related Datasets

| Dataset | Samples | Count Range | Answer Format |
|---------|---------|-------------|---------------|
| StepCountQA-RL-Traj_11_50 | 9,508 | 11–50 | Numeric string |
| StepCountQA-RL-Traj_11_50_NumericOnly | 29,722 | 11–50 | Numeric string |
| **StepCountQA-RL-Traj_11_50_Combined** | **39,230** | **11–50** | **Numeric string** |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("SI-Lab/StepCountQA-RL-Traj_11_50_Combined")
train = ds["train"]
print(train[0])
# {'images': [<PIL.Image>], 'problem': '<image>\nHow many people...', 'answer': '39'}
```

## Dataset Creation

Merged from two sources:

1. **StepCountQA-RL-Traj_11_50** (9,508 rows): Original `<point>` annotation dataset with numeric answers extracted
2. **StepCountQA-RL-Traj_11_50_NumericOnly** (29,722 rows): Converted from 11-50 SFT trajectory dataset

Merge process:
- Extracted image path field from both datasets (streaming, 50 rows/batch)
- Verified zero overlap (0 duplicate images between A and B)
- Streamed A then B in 100-row batches into 4 output shards
- Final row count verified: 9,508 + 29,722 = 39,230 ✓

## License

Please refer to the source dataset license for usage terms.