You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
You agree to use this dataset under the CC-BY-4.0 license with attribution to Datapoint AI, and you agree not to attempt to re-identify annotators. Usage rights for the videos are governed by each model provider's terms.
Log in or Sign Up to review the conditions and access this dataset content.
Text-to-video human preferences: 326K votes across 15 models
This dataset contains the complete voting record behind the Datapoint Video Bench leaderboard: 325,520 validated pairwise votes — exactly 10 for each of 32,552 video pairs. The votes compare 15 text-to-video models on 314 prompts built to stress motion, physics, and temporal consistency, judged by 22,982 annotators in 187 countries. Every generated clip is included as a full-resolution MP4.
Built on the Datapoint annotation platform — purpose-built infrastructure for collecting high-quality human preference data at scale.
Key features
- Ten judges on every pair. Each video pair was judged by 10 people, so every comparison carries a preference margin rather than a single noisy label. The 14 original models meet in a full round-robin on all 313 of their prompts — 3,130 direct votes for each of the 91 pairings. P-Video-2-Pro, added in a second round, meets 13 of them on each of its 313 prompts.
- Prompts that test time, not just frames. 314 prompts across eight categories: camera moves, physics and causality, object permanence, multi-subject interaction, human motion, legible text, stylized animation, and product advertising. Each prompt ships with its category, aspect ratio, and an evaluation rubric.
- Uniform presentation. Every clip is a 1280×720 H.264 MP4 without an audio track, so judges compare what the models generated, not container or sound differences.
- Blind judging. Annotators watched two unlabeled clips with the prompt and one question: "Which video better fulfills the prompt with the highest overall quality? Consider prompt adherence, temporal consistency, motion quality, and visual quality." Model names were never shown.
- Every clip included. The default
videosconfig embeds all 4,708 clips once, with a typedVideocolumn; comparisons join to them by key.
Prompt categories
| Category | Dataset value | Prompts |
|---|---|---|
| Ads & Product | ads_product |
39 |
| Cinematic Camera | cinematic_camera |
40 |
| Human Motion | human_motion |
39 |
| Multi-subject Interaction | multi_subject_interaction |
39 |
| Permanence & State | permanence_state |
40 |
| Physics & Causality | physics_causality |
40 |
| Stylized Animation | stylized_animation |
37 |
| Text in Video | text_in_video |
40 |
| Total | 314 |
Video preview
The same held-out cinematic camera prompt, rendered by three of the models:
The camera pans left to right across a harbor at sunrise, a single smooth pass ending on a red fishing boat; the pan never returns.
Kling 3.0 Pro
Seedance 2.5
Veo 3.1
Dataset structure
| Config | Rows | Description |
|---|---|---|
videos (default) |
4,708 | One MP4 clip per model and prompt, with format metadata |
pairs |
32,552 | One video pair with vote counts, labels, and the winner |
responses |
325,520 | One individual human vote |
prompts |
314 | One prompt with category, aspect ratio, and rubric |
models |
15 | One model with its identifier and display name |
videos
Each row is one model output with the exact MP4 bytes embedded in the video
column, so the Dataset Viewer plays clips directly.
| Column | Type | Description |
|---|---|---|
video |
video | The MP4 clip, decoded by the datasets library on access |
video_key |
string | {model_id}/{prompt_id} join key used by pairs |
prompt_id, prompt_index, category |
string/int | Prompt identity and category |
model_id |
string | Model identifier |
codec, pixel_format, width, height, fps |
mixed | Encoding metadata |
frame_count, duration_seconds, byte_size |
numeric | Clip length and size |
sha256 |
string | SHA-256 digest of the MP4 bytes |
pairs
Each row is one comparison. video_a_key and video_b_key join to videos.
| Column | Type | Description |
|---|---|---|
pair_key |
string | Unique pair identifier |
category, prompt_id |
string | Prompt and its category |
model_a, model_b |
string | The two models being compared |
video_a_key, video_b_key |
string | Join keys into the videos config |
votes_a, votes_b |
int | Vote counts used by the leaderboard; they sum to 10 |
label_a, label_b |
float | Preference fractions. Ties are 0.5/0.5. Use these directly for DPO or reward-model training |
trust_weighted_votes_a, trust_weighted_votes_b |
float | Vote counts weighted by trust score |
winner |
string | a, b, or tie |
num_votes |
int | Total votes for the pair (10) |
responses
Each row is one human vote. Every row counted toward the published leaderboard: the config holds exactly the 10 validated votes per pair that the Elo fit consumed.
| Column | Type | Description |
|---|---|---|
pair_key |
string | Joins to pairs |
prompt_id, category |
string | Prompt and its category |
chosen |
string | a or b, the clip the annotator preferred |
annotator |
string | Salted hash. Stable across the dataset, not linkable to accounts |
trust_score |
float | Annotator trust score (0 to 1) when the vote was cast |
time_taken_ms |
int | Time spent on the judgment. The median is about 11 seconds |
completed_at |
timestamp | When the vote was cast |
country |
string | Annotator country (ISO 3166-1 alpha-2) |
Splits
The videos, pairs, responses, and prompts configs have train and
test splits. The test split holds out 31 prompts, stratified
across the eight categories with a fixed seed. No prompt or clip appears in
both splits, so you can train a reward model on train and evaluate it on
test without prompt leakage. models is a reference table.
Usage
from datasets import load_dataset
pairs = load_dataset("datapointai/text-to-video-human-preferences-326k", "pairs", split="train")
row = pairs[0]
print(row["prompt_id"], row["model_a"], "vs", row["model_b"], "->", row["winner"], f'({row["votes_a"]}-{row["votes_b"]})')
Use for DPO training
With 10 votes per pair, you can select training pairs by preference margin. Filtering to decisive pairs (for example, 7–3 or stronger) removes near-tie label noise:
prompts = {p["prompt_id"]: p["prompt"] for p in load_dataset("datapointai/text-to-video-human-preferences-326k", "prompts", split="train")}
MIN_MARGIN = 0.7 # keep pairs where the winner took >= 7 of 10 votes
def to_dpo(row):
confidence = max(row["label_a"], row["label_b"])
if row["winner"] == "tie" or confidence < MIN_MARGIN:
return None
chosen, rejected = ("a", "b") if row["winner"] == "a" else ("b", "a")
return {
"prompt": prompts[row["prompt_id"]],
"chosen": row[f"video_{chosen}_key"],
"rejected": row[f"video_{rejected}_key"],
"confidence": confidence,
}
dpo_rows = [d for d in (to_dpo(r) for r in pairs) if d]
Get the clips
from datasets import Video
videos = load_dataset("datapointai/text-to-video-human-preferences-326k", "videos", split="train")
videos = videos.cast_column("video", Video(decode=False)) # raw MP4 bytes; decoding needs torchcodec
position = {key: i for i, key in enumerate(videos["video_key"])}
clip = videos[position[dpo_rows[0]["chosen"]]]["video"]["bytes"]
Work with individual votes
# Every vote carries its timing and the annotator's country.
responses = load_dataset("datapointai/text-to-video-human-preferences-326k", "responses", split="train")
considered = responses.filter(lambda r: r["time_taken_ms"] >= 5_000) # votes that took at least 5 seconds
Leaderboard
Elo ratings from a Bradley–Terry fit on raw votes, centered so the 14 original models average 1000. The Overall board gives each of the eight categories an equal share of the fit and matches the live leaderboard. Rank spread shows the ranks each model can hold within its 95% confidence interval.
| Rank | Model | Elo | Rank spread |
|---|---|---|---|
| 1 | Seedance 2.5 | 1049.3 | 1–4 |
| 2 | Gemini Omni 1.1 Flash | 1049.3 | 1–4 |
| 3 | Wan 3.0 | 1045.9 | 1–5 |
| 4 | FLUX.3 Video | 1039.2 | 1–6 |
| 5 | MiniMax H3 Max | 1032.3 | 3–6 |
| 6 | P-Video-2-Pro | 1023.1 | 4–7 |
| 7 | Kling 3.0 Pro | 1010.4 | 6–10 |
| 8 | HappyHorse 1.1 | 998.6 | 7–10 |
| 9 | PixVerse V6 | 997.4 | 7–11 |
| 10 | Ray 3.2 | 997.1 | 7–11 |
| 11 | Grok Imagine 1.5 | 980.3 | 9–12 |
| 12 | LTX 2.5 Pro | 978.4 | 11–13 |
| 13 | Veo 3.1 | 963.8 | 12–14 |
| 14 | Vidu Q3 Pro | 953.5 | 13–14 |
| 15 | Pika 2.5 | 904.3 | 15–15 |
To reproduce the leaderboard, fit a Bradley–Terry model on the responses
config grouped by pair, with raw (unweighted) votes and an equal share for
each category. The trust_weighted_votes_* columns support a sensitivity
analysis: refit with each vote weighted by its trust_score and compare.
Quality and price
Overall Elo plotted against each model's API price per second of generated video, using non-promotional public rates (P-Video-2-Pro is priced at 768p). The highlighted models form the quality–cost frontier: no other model scores higher at the same or lower price.
Comparison to related work
| This dataset | VideoFeedback | Rapidata Pika 2.2 | |
|---|---|---|---|
| Human judgments | 325,520 votes | 34K rated videos | 76K responses |
| Video pairs | 32,552 | — | 1,680 |
| Votes per pair | 10 | — | ~15 per question |
| Models | 15 | 11 | 8 |
| Prompts | 314 | 28,823 | 79 |
Figures for other datasets are counted from their released data files, which can differ from the headline numbers on their cards: VideoFeedback's 37.6K videos include 4,080 real-world videos alongside 33,581 rated generated videos, and Rapidata's card cites about 756K responses while its file holds 76,141.
Intended use
Use this dataset to:
- Train and evaluate reward or preference models for text-to-video generation, including DPO, RLHF, and best-of-N reranking.
- Study inter-annotator agreement and annotation quality at scale.
- Benchmark aggregation methods such as Bradley–Terry variants and trust weighting.
- Audit the published leaderboard.
The clips are one output per model and prompt, collected for evaluation. They are not a curated training corpus for video generation. Don't use this dataset to attempt to identify annotators.
License
Votes, prompts, and all metadata are released under CC-BY-4.0. The videos are outputs of the listed third-party models and are distributed for research and evaluation. Usage rights for model outputs are governed by each provider's terms.
More Datapoint datasets
- text-2-image-human-preferences-2m — 2.16M pairwise votes across 30 text-to-image models, the Datapoint Image Bench voting record
- text-to-speech-human-preferences-315k — 315K pairwise votes across 15 text-to-speech models
- text-2-video-human-preferences-motion-v2-large — 115K pairwise labels on human motion across three quality dimensions
- text-2-video-ranking-human-preferences — 91K ranking labels across 18 text-to-video models
- image-2-video-human-preferences-large — image-to-video preference data
- Full catalog: huggingface.co/datapointai
Citation
@dataset{datapoint_t2v_preferences_2026,
title = {Text-to-Video Human Preferences 326K: the Datapoint Video Bench voting record},
author = {{Datapoint AI}},
year = {2026},
url = {https://huggingface.co/datasets/datapointai/text-to-video-human-preferences-326k},
note = {325,520 pairwise votes over 15 models and 314 prompts, 10 votes per pair}
}
- Downloads last month
- 113