ABRBench Dataset

July 16, 2026 · View on GitHub

ABRBench is a benchmark dataset for Adaptive Bitrate (ABR) research. It reorganizes multiple public network trace collections into two evaluation suites, ABRBench-3G and ABRBench-4G+, with train/test splits and out-of-distribution (OOD) trace sets.

ABRBench is used by SABR to evaluate ABR policy stability, generalization, and robustness across wide-range and unseen network conditions. The related paper is available from IEEE Xplore: SABR: A Stable Adaptive Bitrate Framework Using Behavior Cloning Pretraining and Reinforcement Learning Fine-Tuning.

Video Sets

envivio_3g
- Video chunk sizes sourced from hongzimao/pensieve, identical to the video in hongzimao/video_server.
- Used in ABRBench-3G.
big_buck_bunny
- A customized 40 Mbps cut of Big Buck Bunny.
- Used in ABRBench-4G+.

Trace Dataset Organization and Distribution

ABRBench reorganizes and resplits multiple public trace datasets (note: these are curated versions, not the original official releases).
The train data is aggregated from the provided test sets.

ABRBench-3G

Group	Trace Set	Count	Bandwidth Range (Mbps)	Source
Train	same with test	1828	0.00 – 45.38	-
Test	FCC-16	69	0.00 – 8.95	comyco-lin
	FCC-18	100	0.00 – 41.76	merina
	Oboe	100	0.16 – 9.01	comyco-lin
	Puffer-21	100	0.00 – 25.14	merina
	Puffer-22	100	0.00 – 9.29	merina
OOD	HSR	34	0.00 – 44.68	pitree-dataset

ABRBench-4G+

Group	Trace Set	Count	Bandwidth Range (Mbps)	Source
Train	same with test	262	0.00 – 1890.00	-
Test	Norway 3G	41	0.11 – 7.27	pensieve_retrain
	Lumos 4G	53	0.00 – 270.00	pensieve_retrain
	Lumos 5G	37	0.00 – 1920.00	pensieve_retrain
	Solis Wi-Fi	24	0.00 – 124.00	pensieve_retrain
OOD	Ghent	40	0.00 – 110.97	pitree-dataset
	Lab	61	0.16 – 175.91	pitree-dataset

Dataset Structure

The ABRBench dataset is organized into video files (video/) and network trace files (trace/).
The general trace split principle is:

Train/Test Sets: each trace set contains train/ and test/ subdirectories.
OOD Sets: directly provide trace files, without train/test split.

ABRBench/
├── video/                # Video files
│   ├── big_buck_bunny/   # 4G+ video chunks
│   └── envivio_3g/       # 3G video chunks
│
└── trace/                # Network traces
    ├── ABRBench-3G/      # 3G traces
    │   ├── FCC-16/       # Each trace set contains train / test
    │   │   ├── train/
    │   │   └── test/
    │   ├── FCC-18/
    │   │   ├── train/
    │   │   └── test/
    │   ├── Oboe/
    │   │   ├── train/
    │   │   └── test/
    │   ├── Puffer-21/
    │   │   ├── train/
    │   │   └── test/
    │   ├── Puffer-22/
    │   │   ├── train/
    │   │   └── test/
    │   └── HSR/          # OOD set (raw traces only, no split)
    │
    └── ABRBench-4G+/     # 4G+/5G traces
        ├── Lumos4G/
        │   ├── train/
        │   └── test/
        ├── Lumos5G/
        │   ├── train/
        │   └── test/
        ├── Solis Wi-Fi/
        │   ├── train/
        │   └── test/
        ├── Ghent/        # OOD set (raw traces only, no split)
        └── Lab/          # OOD set (raw traces only, no split)

Citation

If ABRBench or SABR is useful for your research, please cite:

@inproceedings{luo2025sabr,
  title={Sabr: A stable adaptive bitrate framework using behavior cloning pretraining and reinforcement learning fine-tuning},
  author={Luo, Pengcheng and Zhao, Yunyang and Zhang, Bowen and Yang, Genke and Soong, Boon-Hee and Yuen, Chau},
  booktitle={2026 IEEE Wireless Communications and Networking Conference (WCNC)},
  pages={1--6},
  year={2026},
  organization={IEEE}
}