Dataset Card for E-MM1-1M

Dataset Summary

E-MM1-1M is a dataset of 1M+ data groups, pairing data from five modalities: audio, image, video, point cloud, and text. The data and captions are sourced from public data sources.

E-MM1-1M is a subset of E-MM1-100M, which augments the data pairings with human verifications of the connections. The annotations were designed to build pairings between 2 non-text modalities, resulting in triplets of data pairs composed of a text caption, paired modality, and annotated modality. Our pairings include: Text Caption with combinations of (Audio, Points), (Image, Points), (Video, Points) and (Audio, Image).

The dataset was created to advance work on joint embeddings for multimodal applications like cross-modal retrieval.
To visually explore the dataset, please visit our E-MM1 Explorer.

Dataset Structure

The dataset is structured by modalities. We have taken one modality (named annotated_* in the schema below), e.g., image, and it's caption and used that caption to search a big index of, e.g., audio, to find a matching pair for the caption (named paired_* in the schema below). Next, humans have been asked to judge whether the caption from the annotated modality matches the paired modality.
We have sharded annotations from different modalities together:

data
├── audio_to_image.parquet
├── audio_to_points.parquet
├── image_to_audio.parquet
├── image_to_points.parquet
└── video_to_points.parquet

Therefor, you can, e.g., choose to only take the shards where point clouds were annotated against captions:

from datasets import load_dataset
# Full dataset
ds = load_dataset("encord-team/E-MM1-1M", split="train")
# Points annotated
ds = load_dataset("encord-team/E-MM1-1M", split="train", data_files="data/*_points.parquet")
# Images as source
ds = load_dataset("encord-team/E-MM1-1M", split="train", data_files="data/image_*.parquet")

The raw underlying data for this dataset can be downloaded by following the instructions on our GitHub page.

Dataset Splits

We provide two data splits:

• E-MM1-100M (automated): very large, built via nearest-neighbour retrieval for pre-training applications.
• (this dataset) E-MM1-1M (annotated): validated with high quality, human-verified annotations for post-training applications.

The E-MM1-100M split contains the large-scale dataset built with nearest-neighbour retrieval. For each of ~6.7M captions, we retrieved the top-16 nearest neighbours across all modalities, resulting in roughly 1B multimodal connections or 100M groups.

Data Schema

{
  "annotated_dataset_license": "(String) The license of the dataset that the annotated modality belongs to. !! This is not the license of the annotated modality file !!",
  "annotated_end_time": "(Int64) End time of the annotated modality segment in seconds.",
  "annotated_file_id": "(String) The identifier of the annotated modality file from the source dataset.",
  "annotated_file_path": "(String) The path to the annotated modality file if downloaded with download script.",
  "annotated_modality": "(String) The modality of the annotated file.",
  "annotated_source_dataset": "(String) The source dataset of the annotated modality file.",
  "annotated_start_time": "(Int64) Start time of the annotated modality segment in seconds. (if applicable)",
  "annotated_youtube_id": "(String) Youtube ID of the annotated modality file. (if applicable)",
  "annotation": "(Int64) Human annotation of the connection strength between the paired and annotated modalities.",
  "annotation_str": "(String) Human annotation of the connection strength between the paired and annotated modalities as a string.",
  "caption": "(String) Caption sourced with the annotated modality.",
  "encord_annotated_id": "(Int64) Unique ID of the annotated modality file.",
  "encord_paired_id": "(Int64) Unique ID of the paired modality file.",
  "encord_text_id": "(Int64) Unique ID of the text.",
  "paired_dataset_license": "(String) The license of the dataset that the paired modality belongs to. !! This is not the license of the paired modality file !!",
  "paired_end_time": "(Int64) End time of the paired modality segment in seconds. (if applicable)",
  "paired_file_id": "(String) The identifier of the paired modality file from the source dataset.",
  "paired_file_path": "(String) The path to the paired modality file if downloaded with download script.",
  "paired_modality": "(String) The modality of the paired file.",
  "paired_source_dataset": "(String) The source dataset of the paired modality file.",
  "paired_start_time": "(Int64) Start time of the paired modality segment in seconds. (if applicable)",
  "paired_youtube_id": "(String) Youtube ID of the paired modality file. (if applicable)"
}

Additional Information

Usage Documentation

Please find more detailed usage instructions on Github. We have also created an interactive demonstration in Encord for visual exploration of the dataset, where you can find tutorials about the dataset.

Contact

For any questions about the creation of and applications for the dataset, please contact our team at ml@encord.com.

Citation Information

@misc{broadbent2025ebindpracticalapproachspace,
      title={{EBind}: a practical approach to space binding}, 
      author={Jim Broadbent and Felix Cohen and Frederik Hvilshøj and Eric Landau and Eren Sasoglu},
      year={2025},
      eprint={2511.14229},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2511.14229}, 
}