AgPerry/Video-R1-soft-filter

---
dataset_info:
features:
- name: problem_id
dtype: int64
- name: problem
dtype: string
- name: data_type
dtype: string
- name: problem_type
dtype: string
- name: options
dtype: string
- name: solution
dtype: string
- name: path
dtype: string
- name: data_source
dtype: string
splits:
- name: train
num_examples: 160837
size_categories:
- 100K<n<1M
---

# Video-R1-soft-filter

**160,837 samples** filtered from Video-R1-260k using a soft multi-model filter.

## Filtering Logic

Remove a sample **only if 2+ models can answer it text-only** (under circular eval for MCQ).
Keep everything else — this removes the fewest questions while ensuring samples require visual understanding.

| Model | Method | TA count |
|-------|--------|----------|
| GPT-5-mini | Single-pass text-only | 81,361 TA |
| Qwen2.5-VL-7B | Circular eval (MCQ) / pass@10 (non-MCQ) | 95,328 TA |
| Gemini 3.1 Pro | Circular eval (MCQ) / direct (non-MCQ) | 108,808 TA |

A sample is **removed** if ≥2 models can answer it. **Kept** if 0 or 1 models can answer it.

## Statistics

| Metric | Value |
|--------|-------|
| Total samples | 160,837 |
| Reduction from 260k | 38.9% removed |
| Video | 48,475 (30.1%) |
| Image | 112,362 (69.9%) |
| MCQ | 66,578 (41.4%) |
| Non-MCQ | 94,259 (58.6%) |

## Comparison

| Method | Kept | Retention |
|--------|------|-----------|
| GPT single-model (original) | 181,710 | 69.1% |
| **Soft filter (this dataset)** | **160,837** | **61.1%** |
| Triple NTA (3-model all-fail) | 102,057 | 38.8% |
| 2K top-quality | 2,000 | 0.8% |

## Source

Filtered from [Video-R1-260k](https://huggingface.co/datasets/Reacherx/Video-R1) using scripts from the Video-R1 project.