yixin1121/M-Drama

MDrama: Drama Video Understanding Dataset

Dataset Summary

• 32,361 QA annotations over 8,102 short-drama video clips (~52 GB, videos NOT included in this repo).
• Each clip has ~4 annotations: 1 caption/summary + 3 QA (multiple-choice or open-ended).
• Source videos are YouTube short dramas; use the url, start_time, end_time fields to retrieve each clip yourself.
• Pre-decoded frames are also available: the full 2 fps frame package for all clips (train + test) is distributed at yixin1121/M-Drama_Raw — see Raw Frames below if you prefer not to download and slice the YouTube sources yourself.

Fields

Field	Description
qid	Unique question id (train_0 .. train_32360).
video_id	Safe video identifier (video_00000 .. video_08101).
question	Question text (options excluded).
candidates	Options list for MC questions (A. xxx prefixed); null for open-ended.
answer	MC: option letter; caption/summary: reference summary text.
thinking	Reference chain-of-thought reasoning.
caption	Chronological shot-by-shot caption of the clip.
task_type	Task category (e.g., caption, Character-related-Character Identification, ...).
q_type	MC / OE / summary.
url / start_time / end_time	YouTube source of the clip.
stage	Training stage this sample is used in: SFT (all 32,361 samples) or SFT+RL (the 14,135 samples additionally used for RL).

Test Set

• test split: 2,075 QA annotations over 1,036 unique clips (no overlap with train).
• qid: test_0 .. test_2074; video_id: test_video_00000 .. test_video_01035.
• QA only (1,562 MC + 513 OE); thinking is null and stage is "test".
• url / start_time / end_time were manually verified and aligned (as of 2026-09): every clip was checked against its YouTube source, and mis-aligned or dead links were corrected with replacement sources where available. Note that short-drama uploads are frequently taken down: ~9% of the URLs may be dead again at any given time — use the url + timestamps to locate re-uploads when a link fails. Frames for these clips are included in yixin1121/M-Drama_Raw under test/, so the test split can be reproduced even when a link is dead.

Ground-Truth Scene Graphs

gt_graphs.json (~40 MB) provides a ground-truth scene graph for every train clip, keyed by video_id (video_00000 .. video_08101, 8,102 entries):

{
  "video_00000": {
    "nodes": [{"id": "n1", "type": "Scene", "name": "Chinese-style corridor"}, ...],
    "edges": [{"source": "n2", "relation": "AGENT_OF", "target": "n4"}, ...]
  }
}

Graphs were parsed from the reference caption of each clip by Qwen3-14B with the SAGA graph-construction prompt: nodes typed Character / Event / Prop / Scene, semantic edges plus NEXT_EVENT temporal edges. They are the alignment target of SAGA's graph-alignment reward (R_graph = 0.725 * F1_semantic + 0.275 * F1_structural). The test split has no graphs (they are only needed as training reward targets).

Raw Frames (yixin1121/M-Drama_Raw)

Decoded frames for every clip in this dataset — the exact visual input used in the benchmark — are released separately at yixin1121/M-Drama_Raw. Because short-drama uploads are frequently removed, region-locked, or modified, this package guarantees reproducible visual inputs without re-downloading the sources.

Split	Clips	Archives
train/	8,102	train/video_00000.tar … train/video_08101.tar
test/	1,036	test/test_video_00000.tar … test/test_video_01035.tar

• Sampling rate: 2 fps (uniform, uncapped), JPEG, ~190–400 frames per clip (clips are ~1.5–3 min).

• Archive names reuse the same video_id as this dataset, so frames join to annotations directly.

• Each archive is an uncompressed .tar (no gzip/zstd, for fast random seeking); inside, frames sit under a single <video_id>/ directory as 6-digit zero-padded JPEGs in chronological order, plus a done.json with extraction metadata:

  train/video_00000.tar
  └── video_00000/
      ├── 000004.jpg
      ├── 000005.jpg
      ├── ...
      ├── 000242.jpg
      └── done.json   # {"video_id": "video_00000", "frames": 239, "duration_sec": 121.05, "expected_frames": 243}
  

  frames is the number of successfully extracted images and expected_frames is ceil(duration_sec × 2); small deviations in a few clips come from container timestamp offsets or decoding drops.

• Reference evaluation resamples each clip to 192 frames at inference (fps=2.0, max_frames=192); the archives intentionally ship the full uncapped sequences so you can apply your own sampling.

• Clips are identified by code only — no drama titles, source URLs, or original filenames appear in that repo. Use the url / start_time / end_time columns here to map a video_id back to its source.

Usage

import json
from datasets import load_dataset
from huggingface_hub import hf_hub_download

ds = load_dataset("yixin1121/M-Drama")
train = ds["train"]
test = ds["test"]

rl_subset = train.filter(lambda x: x["stage"] == "SFT+RL")

gt_graphs = json.load(open(hf_hub_download(
    "yixin1121/M-Drama", "gt_graphs.json", repo_type="dataset")))
graph = gt_graphs[train[0]["video_id"]]

Loading the raw frames

import io
import tarfile
from PIL import Image
from huggingface_hub import hf_hub_download

row = train[0]
vid = row["video_id"]  # e.g. "video_00000"

path = hf_hub_download("yixin1121/M-Drama_Raw", f"train/{vid}.tar", repo_type="dataset")
with tarfile.open(path) as tar:
    names = sorted(n for n in tar.getnames() if n.endswith(".jpg"))
    frames = [
        Image.open(io.BytesIO(tar.extractfile(n).read())).convert("RGB")
        for n in names
    ]

print(len(frames), row["question"], row["answer"])

hf_hub_download fetches a single archive on demand; for the full set use snapshot_download (the train split is large — prefer per-clip or sharded downloads).

License

The annotations are released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). The frame package at yixin1121/M-Drama_Raw carries the same license and is provided strictly for academic, non-commercial research use.

This dataset only contains annotations (questions, answers, reasoning, and captions). The source videos belong to their respective copyright holders and are NOT distributed with this dataset.