version: 1.0.0 config: REPO_ID: "eduagarcia/open_pt_llm_leaderboard" QUEUE_REPO: eduagarcia-temp/llm_pt_leaderboard_requests RESULTS_REPO: eduagarcia-temp/llm_pt_leaderboard_results RAW_RESULTS_REPO: eduagarcia-temp/llm_pt_leaderboard_raw_results DYNAMIC_INFO_REPO: "eduagarcia-temp/llm_pt_leaderboard_model_info" PATH_TO_COLLECTION: "eduagarcia/portuguese-llm-leaderboard-best-models-65c152c13ab3c67bc4f203a6" IS_PUBLIC: true LEADERBOARD_NAME: "Open PT-LLM Leaderboard" GET_ORIGINAL_HF_LEADERBOARD_EVAL_RESULTS: true TRUST_REMOTE_CODE: true readme: general_description: | 📐 The 🚀 Open PT LLM Leaderboard aims to provide a benchmark for the evaluation of Large Language Models (LLMs) in the Portuguese language across a variety of tasks and datasets. The leaderboard is open to submissions of models from the community and is designed to be a resource for researchers, practitioners, and enthusiasts interested in the development and evaluation of LLMs for the Portuguese language. If you have any questions, suggestions, or would like to contribute to the leaderboard, please feel free to reach out at [@eduagarcia](https://linktr.ee/eduagarcia). support_description: | This leaderboard is made possible by the support of the [Center of Excelence in AI (CEIA)](https://ceia.ufg.br/) at the [Federal University of Goiás (UFG)](https://international.ufg.br/). about_description: | The 🚀 Open PT-LLM Leaderboard is a benchmark for the evaluation of Large Language Models (LLMs) in the Portuguese language. The leaderboard is open to submissions of models from the community and is designed to be a resource for researchers, practitioners, and enthusiasts interested in the development and evaluation of LLMs for the Portuguese language. Supported by the [Center of Excelence in AI (CEIA)](https://ceia.ufg.br/) at the [Federal University of Goiás (UFG)](https://international.ufg.br/), this leaderboard operates on a backend of Nvidia A100-80G GPUs. Evaluations are subject to resource availability, which is not exclusive. Therefore, please be patient if your model is in the queue. If you'd like to support the leaderboard, feel free to reach out. This is a fork of the 🤗 Open LLM Leaderboard with portuguese benchmarks. tasks: enem_challenge: benchmark: enem_challenge col_name: ENEM task_list: - enem_challenge metric: acc few_shot: 3 limit: null baseline: 20.0 #random baseline #https://www.sejalguem.com/enem #https://vestibular.brasilescola.uol.com.br/enem/confira-as-medias-e-notas-maximas-e-minimas-do-enem-2020/349732.html human_baseline: 35.0 # ~60 / 180 acertos - nota ~500 expert_human_baseline: 70.0 # ~124 / 180 acertos - nota ~700 description: "The Exame Nacional do Ensino Médio (ENEM) is an advanced High-School level exam widely applied every year by the Brazilian government to students that wish to undertake a University degree. This dataset contains 1,430 questions that don't require image understanding of the exams from 2010 to 2018, 2022 and 2023." link: https://huggingface.co/datasets/eduagarcia/enem_challenge sources: ["https://www.ime.usp.br/~ddm/project/enem/", "https://github.com/piresramon/gpt-4-enem", "https://huggingface.co/datasets/maritaca-ai/enem"] baseline_sources: ["https://www.sejalguem.com/enem", "https://vestibular.brasilescola.uol.com.br/enem/confira-as-medias-e-notas-maximas-e-minimas-do-enem-2020/349732.html"] bluex: benchmark: bluex col_name: BLUEX task_list: - bluex metric: acc few_shot: 3 limit: null baseline: 22.5 #random baseline #https://www.comvest.unicamp.br/wp-content/uploads/2023/08/Relatorio_F1_2023.pdf 56% mean - 88% @ top-.99 #https://acervo.fuvest.br/fuvest/2018/FUVEST_2018_indice_discriminacao_1_fase_ins.pdf 43,4% - ~77% @ top-.99 human_baseline: 50.0 expert_human_baseline: 82.5 description: "BLUEX is a multimodal dataset consisting of the two leading university entrance exams conducted in Brazil: Convest (Unicamp) and Fuvest (USP), spanning from 2018 to 2024. The benchmark comprises of 724 questions that do not have accompanying images" link: https://huggingface.co/datasets/eduagarcia-temp/BLUEX_without_images sources: ["https://github.com/portuguese-benchmark-datasets/bluex", "https://huggingface.co/datasets/portuguese-benchmark-datasets/BLUEX"] baseline_sources: ["https://www.comvest.unicamp.br/wp-content/uploads/2023/08/Relatorio_F1_2023.pdf", "https://acervo.fuvest.br/fuvest/2018/FUVEST_2018_indice_discriminacao_1_fase_ins.pdf"] oab_exams: benchmark: oab_exams col_name: OAB Exams task_list: - oab_exams metric: acc few_shot: 3 limit: null baseline: 25.0 #random baseline #https://fgvprojetos.fgv.br/publicacao/exame-de-ordem-em-numeros # 46% # http://fgvprojetos.fgv.br/publicacao/exame-de-ordem-em-numeros-vol3 # Acertou +70% = 17214 / 638500 = top-97,5% # desvio top-97,5% -> 46 - 70.0% = 24 # z score 97,5% ~ 1,9675 # desvio padrao estimado -> 12,2 # top 99% = 46 + 2,33*12,2 = ~75.0 human_baseline: 46.0 expert_human_baseline: 75.0 description: OAB Exams is a dataset of more than 2,000 questions from the Brazilian Bar Association's exams, from 2010 to 2018. link: https://huggingface.co/datasets/eduagarcia/oab_exams sources: ["https://github.com/legal-nlp/oab-exams"] baseline_sources: ["http://fgvprojetos.fgv.br/publicacao/exame-de-ordem-em-numeros", "http://fgvprojetos.fgv.br/publicacao/exame-de-ordem-em-numeros-vol2", "http://fgvprojetos.fgv.br/publicacao/exame-de-ordem-em-numeros-vol3"] assin2_rte: benchmark: assin2_rte col_name: ASSIN2 RTE task_list: - assin2_rte metric: f1_macro few_shot: 15 limit: null baseline: 50.0 #random baseline human_baseline: null expert_human_baseline: null description: "ASSIN 2 (Avaliação de Similaridade Semântica e Inferência Textual - Evaluating Semantic Similarity and Textual Entailment) is the second edition of ASSIN, an evaluation shared task in the scope of the computational processing of Portuguese. Recognising Textual Entailment (RTE), also called Natural Language Inference (NLI), is the task of predicting if a given text (premise) entails (implies) in other text (hypothesis)." link: https://huggingface.co/datasets/eduagarcia/portuguese_benchmark sources: ["https://sites.google.com/view/assin2/", "https://huggingface.co/datasets/assin2"] assin2_sts: benchmark: assin2_sts col_name: ASSIN2 STS task_list: - assin2_sts metric: pearson few_shot: 15 limit: null baseline: 0.0 #random baseline human_baseline: null expert_human_baseline: null description: "Same as dataset as above. Semantic Textual Similarity (STS) ‘measures the degree of semantic equivalence between two sentences’." link: https://huggingface.co/datasets/eduagarcia/portuguese_benchmark sources: ["https://sites.google.com/view/assin2/", "https://huggingface.co/datasets/assin2"] faquad_nli: benchmark: faquad_nli col_name: FAQUAD NLI task_list: - faquad_nli metric: f1_macro few_shot: 15 limit: null baseline: 45.6 #random baseline human_baseline: null expert_human_baseline: null description: "FaQuAD is a Portuguese reading comprehension dataset that follows the format of the Stanford Question Answering Dataset (SQuAD). The dataset aims to address the problem of abundant questions sent by academics whose answers are found in available institutional documents in the Brazilian higher education system. It consists of 900 questions about 249 reading passages taken from 18 official documents of a computer science college from a Brazilian federal university and 21 Wikipedia articles related to the Brazilian higher education system. FaQuAD-NLI is a modified version of the FaQuAD dataset that repurposes the question answering task as a textual entailment task between a question and its possible answers." link: https://huggingface.co/datasets/ruanchaves/faquad-nli sources: ["https://github.com/liafacom/faquad/"] sparrow_pt: benchmark: sparrow_pt col_name: Sparrow POR task_list: - sparrow_emotion-2021-cortiz-por - sparrow_hate-2019-fortuna-por - sparrow_sentiment-2016-mozetic-por - sparrow_sentiment-2018-brum-por metric: f1_macro few_shot: 25 limit: 500 baseline: 29.5 #random baseline [3.3, 48.8, 33.1, 33.0] human_baseline: null expert_human_baseline: null description: "SPARROW is a multilingual evaluation benchmark for sociopragmatic meaning understanding. SPARROW comprises 169 datasets encompassing 64 different languages, this split evaluates only on the validation set of 4 datasets avaliable for the Portuguese language. One on hate speech detection by Fortuna et al. (2019), one on emotion detection by Cortiz et al. (2021) and two on sentiment analysis by Mozetic et al. (2016) and Brum et al. (2018). All were extracted and manually annotated from Twitter/X." link: https://huggingface.co/datasets/UBC-NLP/sparrow sources: ["https://sparrow.dlnlp.ai/", "https://aclanthology.org/W19-3510/", "https://arxiv.org/abs/2108.07638", "https://aclanthology.org/L18-1658/", "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0155036"]