Card: refresh body to current state (93 models, 22 tasks/7 categories, mean_22, arXiv citation)

README.md

@@ -97,38 +97,39 @@ models:
 
 # MTEB-BR — Brazilian Portuguese Massive Text Embedding Benchmark
 
-A Massive Text Embedding Benchmark for **Brazilian Portuguese**.
+A Massive Text Embedding Benchmark for **Brazilian Portuguese** — native, not translated.
 
-- **56 models** evaluated on **16 native PT-BR tasks** (no machine-translated mMARCO/mkqa)
-- Tasks: classification, NLI, STS, clustering, retrieval (incl. Quati Pool), reranking
-- Headline metric: **mean_16** (average across all 16 tasks)
-- All raw results: [MTEB-BR/mteb-pt-results](https://huggingface.co/datasets/MTEB-BR/mteb-pt-results)
-- Source code: [github.com/tardellirs/mteb-br](https://github.com/tardellirs/mteb-br)
+- **93 models** (73 open-weight + 20 closed commercial APIs) evaluated on **22 native Brazilian-Portuguese tasks** across **7 categories**
+- Admits only data created or found in Portuguese; machine-translated benchmarks (e.g. mMARCO, mkqa) are excluded by construction
+- Headline metric: **mean_22** (average across all 22 tasks), reported with per-task bootstrap confidence intervals, paired-bootstrap significance, and IRT task discrimination
+- Website: [mteb-br.org](https://mteb-br.org) · Paper: [arXiv:2607.04581](https://arxiv.org/abs/2607.04581)
+- All raw results: [MTEB-BR/mteb-pt-results](https://huggingface.co/datasets/MTEB-BR/mteb-pt-results) · Source code: [github.com/tardellirs/mteb-br](https://github.com/tardellirs/mteb-br)
 
-## Tasks (16)
+## Tasks (22)
 
 | Category | Tasks |
 |---|---|
-| Classification | HateBR, OffComBR, TweetSentBR, ToxSynPT |
-| Pair classification (NLI) | AssinRTE, InferBR |
-| STS | AssinSTS |
-| Clustering | MedPTClustering, WikipediaPTCategoriesClusteringP2P |
-| Retrieval | Quati (Pool), JurisTCU, BRTaxQAR, FaQuADIR, MedPTRetrieval |
-| Reranking | QuatiReranking, JurisTCUReranking |
+| Classification | HateBR, ToxSyn-PT, FactckBR, PortuLexRRIP |
+| Multi-label classification | BrighterEmotion |
+| Pair classification | ASSIN-RTE, InferBR |
+| Semantic textual similarity (STS) | ASSIN-STS, ASSIN2-STS |
+| Clustering | WikipediaPT-Categories, MedPT, JurisTCU, SciELO, StackOverflow-PT |
+| Retrieval | Quati, JurisTCU, BRTaxQA, FaQuAD, MedPT, FaQ-Bacen |
+| Reranking | Quati, JurisTCU |
 
 ## How to add a model
 
 Submit via GitHub Issues at [tardellirs/mteb-br/issues](https://github.com/tardellirs/mteb-br/issues)
-with model_id, eval JSONs, and reproduction script.
+with the model ID, evaluation JSONs, and a reproduction script.
 
 ## Citation
 
 ```bibtex
-@misc{mteb-por-2026,
-  title  = {MTEB-BR: The Brazilian Portuguese MTEB Sub-Benchmark},
-  author = {Tardelli, R.S.},
-  year   = {2026},
-  url    = {https://huggingface.co/spaces/MTEB-BR/leaderboard}
+@article{stekel2026mtebbr,
+  title   = {MTEB-BR: A Text Embedding Benchmark for Brazilian Portuguese},
+  author  = {Stekel, Tardelli Ronan Coelho},
+  journal = {arXiv preprint arXiv:2607.04581},
+  year    = {2026}
 }
 ```