You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

MOEB results: omni embedding models on mteb

Raw mteb result files for 11 omni (text, image, audio, video) embedding models. The task set is the one in code/run_moeb_omni_2.py: MMTEB v3 (MTEB(Multilingual, v2) without 5 tasks), every cross-modal task in mteb, and the single-modality tasks of MIEB(lite), MAEB(v1) and MVEB(v1) with their variants. That is 392 tasks per model.

Status: incomplete. Uploaded 2026-11-04. About 2,600 of 3,920 model–task results are done. The run stopped when the compute project ran out of GPU time.

Contents

Path What
results/<model>/<revision>/<task>.json One mteb TaskResult per model and task
summary.json, summary.txt Mean main_score per model and benchmark, done/expected, missing tasks. Snapshot of 2026-11-04 01:07 UTC; results finished after that time are in results/ but not in the summary. Rebuild it with code/summarize_moeb.py.
code/ Runner, Slurm scripts, summary script, and run notes
code/mteb_local_changes.patch Local fixes on top of mteb commit e9c128f2c331eb271c18b442734940549f29edc2

Models

nvidia/omni-embed-nemotron-3b, EximiusLabs/fusion-embedding-1-2b-preview, EximiusLabs/fusion-embedding-2-2b-preview, encord-team/ebind-full, encord-team/ebind-audio-vision, jinaai/jina-embeddings-v5-omni-small, jinaai/jina-embeddings-v5-omni-nano, LanguageBind/LanguageBind_Omni, google/embeddinggemma-2, BidirLM/BidirLM-Omni-2.5B-Embedding.

vidore/colqwen-omni-v0.1 was dropped part-way. Its partial results are included, but the summary leaves it out. It is a multi-vector model, so classification, clustering and pair-classification tasks do not apply to it.

Caveats

  • BidirLM-Omni video scores change with mteb PR #5652, which was not merged at run time.
  • The local patch changes results for three cases:
    • LanguageBind: grayscale and RGBA images are converted to RGB.
    • omni-embed-nemotron: a video time value is passed as a tensor; without the fix every video task fails.
    • Multilabel classification: a dataset index fix for datasets 4.0.
  • Some datasets are gated on the Hub (for example facebook/winoground). Tasks on those datasets are missing.
Downloads last month
27