ByteDance/LiveStreamingRiskControl

Live or Lie — Live Streaming Room Risk Assessment (May/June 2025)

This dataset contains live-streaming room interaction logs for room-level risk assessment under weak supervision. It is the official dataset for the research presented in the papers:

• Outsmarting the Chameleon: Counterfactual Decoupling for Tactical OOD Shifts in Live Streaming Risk Assessment (Hugging Face Papers)
• Live or Lie: Action-Aware Capsule Multiple Instance Learning for Risk Assessment in Live Streaming Platforms (arXiv)

Project Page: https://qiaoyran.github.io/LiveStreamingRiskAssessment/ GitHub: https://github.com/bytedance/AC-MIL

Dataset Summary

Each example corresponds to a single live-streaming room and is labeled as risky (> 0) or normal (= 0).

The task is designed for early detection: each room’s action sequence is truncated to the first 30 minutes, and can be structured into user–timeslot capsules for models such as AC-MIL or LPCD.

File Structure

The dataset is organized into two time-indexed subsets (May and June). Large LMDB data files are provided in multiple .part chunks to comply with storage limits.

.
├── final_May_hard1_masked_encoded.lmdb/
│   ├── data.mdb.00.part
│   ├── data.mdb.01.part
│   ├── data.mdb.02.part
│   ├── data.mdb.03.part
│   └── lock.mdb
├── final_June_hard1_masked_encoded.lmdb/
│   ├── data.mdb.00.part
│   ├── data.mdb.01.part
│   └── lock.mdb
├── May_train.csv
├── May_val.csv
├── May_test.csv
├── June_train.csv
├── June_val.csv
└── June_test.csv

Languages

• Predominantly Chinese (zh): user behaviors are presented in Chinese, e.g., "主播口播：...". These action descriptions are encoded as action vectors via a Chinese-BERT model.

Data Structure

Each room has a label and a sequence of actions:

• room_id (string)
• label (int32, {0,1,2,3}))
• patch_list (list of tuples):
  • u_idx (string): user identifier within a room
  • t (int32): time index along the room timeline
  • l (int32): capsule index
  • action_id (int32): action type ID
  • action_vec (list<float16> or null): action features encoded from masked action descriptions
  • timestamp (string): action timestamp
  • action_desc (string): textual action descriptions
  • user_id (string): user identifier across rooms

Action Space

The setup includes both viewer interactions (e.g., room entry, comments, likes, gifts, shares, etc.) and streamer activities (e.g., start stream, speech transcripts via voice-to-text, OCR-based visual content monitoring). Text-like fields are discretized as part of platform inspection/sampling.

Data Splits

Split	#Rooms	Avg. actions	Avg. users	Avg. time (min)
May train	176,354	709	35	30.0
May val	23,859	704	36	29.6
May test	22,804	740	37	29.7
June train	80,472	700	36	30.0
June val	10,934	767	40	29.1
June test	11,116	725	37	29.1

Quickstart

1. Reconstruct the LMDB files

Before loading the data, you must merge the split parts back into a single data.mdb file for each subset.

# Reconstruct May Dataset
cd final_May_hard1_masked_encoded.lmdb
cat data.mdb.*.part > data.mdb
cd ..

# Reconstruct June Dataset
cd final_June_hard1_masked_encoded.lmdb
cat data.mdb.*.part > data.mdb
cd ..

2. Loading Data

Install the Python package: pip install lmdb

import lmdb
import pickle

# Example: read a specific room
room_id = 0 
lmdb_path = "final_May_hard1_masked_encoded.lmdb"

env = lmdb.open(
    lmdb_path,
    readonly=True,
    lock=False,
    map_size=240 * 1024 * 1024 * 1024,
    readahead=False,
)

with env.begin() as txn:
    value = txn.get(str(room_id).encode())
    if value is not None:
        data = pickle.loads(value)
        patch_list = data["patch_list"] # list of tuples
        room_label = data["label"]

env.close()

Security and Privacy

To ensure the security and privacy of users, all data collected from live rooms has been anonymized and masked, preventing any content from being linked to a specific individual. In addition, action vectors are re-encoded from the masked action descriptions.

Considerations for Using the Data

Intended Use

• Research on weakly-supervised risk detection / MIL in live streaming.
• Early-warning room-level moderation signals.
• Interpretability over localized behavior segments.

Out-of-scope / Misuse

• Do not use this dataset to identify, profile, or target individuals.
• Do not treat predictions as definitive enforcement decisions without human review.

Bias and Limitations

• Sampling bias: negatives are downsampled (1:10); reported metrics and thresholds should account for this.
• Domain specificity: behavior patterns are platform- and policy-specific; transfer to other platforms may be limited.

License

This dataset is licensed under CC BY 4.0.

Citation

@article{qiao2026live,
  title={Live or Lie: Action-Aware Capsule Multiple Instance Learning for Risk Assessment in Live Streaming Platforms},
  author={Qiao, Yiran and Chen, Jing and Ao, Xiang and Zhong, Qiwei and Liu, Yang and He, Qing},
  journal={arXiv preprint arXiv:2602.03520},
  year={2026}
}