chowmeinj/streamingomni_datasets

StreamingOmni-Combined

Dataset Description

This dataset combines three StreamingOmni datasets:

• ASR: Automatic Speech Recognition data from LibriSpeech, TEDLIUM, and VoxPopuli
• MT: Machine Translation data from IWSLT 2017 (en-fr, en-de)
• QA: Question Answering data from SQuAD

The dataset is formatted in the StreamingOmni unified format, designed for training multi-modal Omni models.

Data Splits

• Train: ~171,506 samples
• Validation: ~21,653 samples
• Test: ~31,548 samples

Capabilities

This dataset supports:

1. Automatic Speech Recognition (ASR): Audio -> Text
2. Machine Translation (MT): Text (source) -> Text (target)
3. Question Answering (QA): Question + Context -> Answer

Source Datasets

• ASR: LibriSpeech (CC-BY-4.0), TEDLIUM, VoxPopuli (CC0-1.0)
• MT: IWSLT 2017 (en-fr, en-de)
• QA: SQuAD (CC-BY-4.0)

License

CC BY-NC 4.0

Usage Example

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("your-username/streamingomni-combined")

# Access a sample
sample = dataset["train"][0]
print(f"Task: {sample[\"task_type\"]}")
print(f"Instruction: {sample[\"instruction\"]}")
print(f"Target: {sample[\"targets\"][\"answer\"]}")