---
license: cc-by-nc-4.0
task_categories:
  - text-to-video
tags:
  - video
  - video-pair
  - reward-model
  - dpo
  - t2v
  - broad-pair
pretty_name: DenseDPO T2V Broad-Pair Dataset (v2)
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
dataset_info:
  features:
    - name: pair_id
      dtype: string
    - name: tier
      dtype: string
    - name: model_a
      dtype: string
    - name: model_b
      dtype: string
    - name: video_a
      dtype:
        video: {}
    - name: video_b
      dtype:
        video: {}
    - name: stem
      dtype: string
    - name: prompt
      dtype: string
    - name: task
      dtype: string
  splits:
    - name: train
      num_examples: 6519
---

# DenseDPO T2V Broad-Pair Dataset (v2)

Cross-model **text-to-video (T2V)** generation pairs for training video reward
models (RM) and DPO-style preference learning.

The HF Dataset Viewer renders each row as **prompt + two videos side-by-side**.

## Generation task

All videos are generated **T2V** from a shared text prompt. For every pair,
both videos share the same prompt, so the primary comparison axis is the model
identity itself.

## Plan-A tier structure

Models are grouped into two resolution tiers; pairs only form **within** a tier
to keep resolution as a constant signal for the RM.

| Tier | Models | Resolution | Frames @ FPS |
|------|--------|------------|--------------|
| 480p (low-res)  | Helios-Distilled, LongCat-Video | 384×640 / 480×832 | 99@16 / 93@15 |
| 720p (high-res) | LTX-Video-2.3-distilled, SkyReels-V2, Wan-2.2 | 704×1280 / 720×1280 | 121@24 |

Pair-types (one prompt → exactly one pair-type):
- 480p: `(helios, longcat)`
- 720p: `(ltx, skyreels)`, `(ltx, wan)`, `(skyreels, wan)`

## File layout

```
t2v_data_v2_hub/
├── README.md
├── data/
│   └── train-*.parquet
└── pairs_summary.json
```

## Schema

| Column | Type | Notes |
|---|---|---|
| `pair_id` | string | unique pair id |
| `tier` | string | `"480p"` or `"720p"` |
| `model_a`, `model_b` | string | alphabetically sorted within a pair |
| `video_a`, `video_b` | **Video()** | mp4 bytes embedded in parquet; renders inline in HF Viewer |
| `stem` | string | filename-derived prompt stem |
| `prompt` | string | shared text caption from the manifest |
| `task` | string | always `"T2V"` |

## Loading

```python
from datasets import load_dataset

ds = load_dataset("<hf-user>/t2v_data_v2", split="train")
print(ds[0]["prompt"])
print(ds[0]["video_a"])   # decord/pyav-decoded Video object
print(ds[0]["video_b"])
```

When loading from disk:

```python
from datasets import load_dataset
ds = load_dataset("parquet", data_files="t2v_data_v2_hub/data/train-*.parquet",
                  split="train")
```

Video bytes are embedded directly in the parquet shards, so no `videos/` folder is required.

## Counts

See `pairs_summary.json` for the exact per-pair-type breakdown.

### This build

- `(helios, longcat)` [480p]: **880** pairs
- `(ltx, skyreels)` [720p]: **800** pairs
- `(ltx, wan)` [720p]: **800** pairs
- `(skyreels, wan)` [720p]: **800** pairs

- Dropped-source pairs: **3280**
- Selected T2V pairs: **3239**
- Total pairs: **6519**
- Orphan stems (1 model only, skipped): 0
- Cross-tier pairs (skipped): 0
- Stems with 3+ models (skipped): 0


## Pushing to HF Hub

```bash
hf upload-large-folder <hf-user>/t2v_data_v2 t2v_data_v2_hub/ --repo-type=dataset
```


## Selected T2V Source

Selected T2V pairs from selected_t2v_filtered.json are included in the same train split with `task` set to `T2V`. Video bytes are embedded in 12 parquet shards; no `videos/` folder is required.