Salesforce/ST-Evidence-Bench

ST-Evidence Benchmark Dataset

ST-Evidence is a comprehensive benchmark for evaluating Spatial-Temporal Evidence generation in video understanding. It contains two tasks: Generation (Gen) and Multiple Choice Question (MCQ).

This was released for research purposes only, in support of the academic paper Evidence-Backed Video Question Answering.

Dataset Overview

• Total Videos: ~1,300 videos at 6fps
• Annotations: Question-Answer pairs with temporal segments and spatial masks
• Tasks: Generation and MCQ
• Domains: Diverse video content

Files Structure

ST-Evidence/
├── st_evidence_gen/              # Generation Task
│   ├── st_evidence_gen.csv       # 924KB - Annotations (entry_id, question, answer, segments, etc.)
│   ├── videos_6fps.tar.gz        # 8.3GB - Video files at 6fps
│   └── masks.tar.gz              # 560MB - Ground truth spatial masks
│
└── st_evidence_mcq/              # Multiple Choice Question Task
    ├── st_evidence_mcq.csv       # 313KB - MCQ annotations
    ├── mask_options.json         # 575KB - Mask options metadata
    ├── temp_options.json         # 679KB - Temporal options metadata
    └── options.tar.gz            # 1.5GB - Pre-rendered option masks (1,298 entries)

Generation Task (st_evidence_gen)

Data Format

st_evidence_gen.csv contains the following columns:

• entry_id: Unique identifier for each question
• video_id: Video identifier
• video_path: Relative path to video file
• question: Question text
• candidates: List of answer options (for reference)
• answer: Ground truth answer
• segment: Temporal evidence segments [[start1, end1], [start2, end2], ...]

Usage

import pandas as pd
import tarfile

# Load annotations
df = pd.read_csv('st_evidence_gen.csv')

# Extract videos
with tarfile.open('videos_6fps.tar.gz', 'r:gz') as tar:
    tar.extractall('videos_6fps/')

# Extract ground truth masks
with tarfile.open('masks.tar.gz', 'r:gz') as tar:
    tar.extractall('masks/')

Evaluation Metrics

• QA Accuracy: Percentage of correct answers
• Temporal IoU: Intersection over Union for temporal segments
  • mIoU, TIoU@0.3, TIoU@0.5
• Temporal IoP: Intersection over Prediction
  • mIoP, TIoP@0.3, TIoP@0.5
• Spatial Quality (if masks generated):
  • J score (Jaccard/IoU)
  • F score (contour-based)
  • J&F score (average)

Distribution

• clevrer: 366 (13.53%)
• pt: 782 (28.90%)
• qa_ego4d_trimmed: 270 (9.98%)
• star: 327 (12.08%)
• vicas: 961 (35.51%)

MCQ Task (st_evidence_mcq)

Data Format

st_evidence_mcq.csv contains:

• entry_id: Unique identifier
• video_id: Video identifier
• video_path: Path to video
• question: Question text
• candidates: Answer options
• answer: Correct answer
• segment: Temporal evidence
• mask_options: Reference to mask options
• temp_options: Reference to temporal options

mask_options.json: Contains spatial mask options for each question temp_options.json: Contains temporal segment options for each question options.tar.gz: Pre-rendered mask visualizations for options (1,298 entries)

Usage

import json
import pandas as pd

# Load MCQ annotations
df = pd.read_csv('st_evidence_mcq.csv')

# Load options
with open('mask_options.json', 'r') as f:
    mask_options = json.load(f)

with open('temp_options.json', 'r') as f:
    temp_options = json.load(f)

# Extract option masks
with tarfile.open('options.tar.gz', 'r:gz') as tar:
    tar.extractall('options/')

Evaluation Metrics

Same as Generation task, but with multiple-choice format.

Download & Setup

Using HuggingFace Hub

from huggingface_hub import snapshot_download

# Download entire dataset
snapshot_download(
    repo_id="Salesforce/ST-Evidence",
    repo_type="dataset",
    local_dir="./st_evidence_data"
)

Manual Download

1. Download all files from this repository
2. Extract compressed files:

   tar -xzf videos_6fps.tar.gz
   tar -xzf masks.tar.gz
   tar -xzf options.tar.gz
   

Citation

If you use this dataset, please cite:

@article{st-evidence2025,
  title={ST-Evidence: A Benchmark for Spatial-Temporal Evidence in Video Understanding},
  author={Wang, Shijie and others},
  year={2025}
}

License

CC-BY-NC 4.0

Version

• Version: 1.0
• Release Date: 2026-07-TBD
• Total Size: ~10.4 GB (compressed)