README.md

metadata

license: unknown
task_categories:
  - video-text-to-text
  - visual-question-answering
language:
  - en
tags:
  - video
  - captions
  - embeddings
  - LVBench
  - VideoMME
  - MLVU
  - LongVideoBench
pretty_name: Merged Video Benchmark Captions and Semantic Embeddings

Merged Video Benchmark Captions and Semantic Embeddings

This dataset contains merged semantic caption segments and aligned text embeddings for four video benchmarks from a local DynamicvideoAgentRL/LVU cache. It includes captions and embeddings only, not source videos.

Files

• captions.parquet: one row per caption segment.
• semantic_vectors.float32.npy: NumPy array with shape 244909 x 3072; row i matches captions.parquet row where row_id == i.
• semantic_norms.float32.npy: L2 norms aligned by row_id.
• summary.json: generation metadata and counts.

Usage

import numpy as np
import pandas as pd

df = pd.read_parquet('captions.parquet')
vecs = np.load('semantic_vectors.float32.npy', mmap_mode='r')
row = df.iloc[0]
embedding = vecs[row.row_id]

Schema

• row_id: row index into semantic_vectors.float32.npy and semantic_norms.float32.npy.
• benchmark: one of LVBench, videomme, mlvu, LongVideoBench.
• video_id: video id under that benchmark.
• video_key: joined key <benchmark>/<video_id>.
• doc_id: original semantic caption key, usually <start>_<end>.
• start_sec, end_sec: segment window in seconds.
• caption: merged caption text used by semantic retrieval.
• clip_caption: visual caption component when available.
• ocr_text: OCR/subtitle text component when available.
• entities_json: JSON-encoded entities list.
• source_json: JSON-encoded source flags.
• source_path: original relative path under indexes/semantic.
• embedding_model: embedding model recorded in source metadata.
• embedding_dim: embedding dimension.

Counts

{
  "LVBench": {
    "videos": 103,
    "segments": 26048
  },
  "videomme": {
    "videos": 900,
    "segments": 57893
  },
  "mlvu": {
    "videos": 1991,
    "segments": 138184
  },
  "LongVideoBench": {
    "videos": 753,
    "segments": 22784
  }
}

Total videos: 3747
Total segments: 244909
Embedding model(s): text-embedding-3-large