---
title: MTEB Portuguese
emoji: 🏆
colorFrom: green
colorTo: yellow
sdk: static
pinned: false
license: apache-2.0
---

# MTEB Portuguese

A public benchmark for evaluating text embedding models on **Brazilian Portuguese**, built as a thin extension on top of the [`mteb`](https://github.com/embeddings-benchmark/mteb) library.

## What you'll find here

- 🏆 **[Leaderboard](https://huggingface.co/spaces/mteb-pt/leaderboard)** — interactive ranking, 54 models × 16 tasks, Pareto chart
- 📊 **[`mteb-pt-results`](https://huggingface.co/datasets/mteb-pt/mteb-pt-results)** — all per-task JSONs + per-query parquets, ~1100 files
- 💻 **[GitHub repo](https://github.com/tardellirs/mteb-pt)** — task definitions, evaluation scripts, paper sources, issue templates

## Submit a model

We accept submissions via either channel — pick whichever fits:

- 💬 [HF Discussion on the results dataset](https://huggingface.co/datasets/mteb-pt/mteb-pt-results/discussions/new)
- 🐛 [GitHub Issue with the model template](https://github.com/tardellirs/mteb-pt/issues/new?template=submit-model.yml)

Required for a submission:
1. `model_id` (HF repo path or vendor product name)
2. Per-task result JSONs for the 16 headline tasks
3. Reproducible evaluation command

We re-run a sample of each submission to verify before merging.

## Propose a new task

Open a [GitHub Issue with the task template](https://github.com/tardellirs/mteb-pt/issues/new?template=propose-task.yml) describing the dataset, license, size, and discrimination evidence. A task is accepted if it's native PT-BR (not machine-translated), has clear licensing, and discriminates between models.

## Maintainer

**Tardelli Stekel** — IFSP, São Paulo, Brazil
✉️ <stekel@ifsp.edu.br>

Contributions, corrections, and discussion all welcome.

## Citation

```bibtex
@misc{mteb-portuguese-2026,
  title  = {MTEB Portuguese: A Massive Text Embedding Benchmark for Brazilian Portuguese},
  author = {Stekel, Tardelli},
  year   = {2026},
  url    = {https://huggingface.co/spaces/mteb-pt/leaderboard}
}
```

## Acknowledgments

Built on top of the [`mteb`](https://github.com/embeddings-benchmark/mteb) library by Enevoldsen et al. (2025). Task datasets contributed by their original authors. Compute provided by Modal.