tasks_config/pt_config.yaml

version: 1.1.0
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 Portuguese LLM Leaderboard"
  GET_ORIGINAL_HF_LEADERBOARD_EVAL_RESULTS: true
  TRUST_REMOTE_CODE: true
  SHOW_INCOMPLETE_EVALS: false
  REQUIRE_MODEL_CARD: true
  REQUIRE_MODEL_LICENSE: false
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 <a href="https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard" target="_blank">🤗 Open LLM Leaderboard</a> 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://www.ime.usp.br/~ddm/project/enem/ENEM-GuidingTest.pdf
    sources: ["https://huggingface.co/datasets/eduagarcia/enem_challenge", "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://arxiv.org/abs/2307.05410
    sources: ["https://huggingface.co/datasets/eduagarcia-temp/BLUEX_without_images", "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://arxiv.org/abs/1712.05128
    sources: ["https://huggingface.co/datasets/eduagarcia/oab_exams", "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://dl.acm.org/doi/abs/10.1007/978-3-030-41505-1_39
    sources: ["https://huggingface.co/datasets/eduagarcia/portuguese_benchmark", "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://dl.acm.org/doi/abs/10.1007/978-3-030-41505-1_39
    sources: ["https://huggingface.co/datasets/eduagarcia/portuguese_benchmark", "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://ieeexplore.ieee.org/abstract/document/8923668
    sources: ["https://github.com/liafacom/faquad/", "https://huggingface.co/datasets/ruanchaves/faquad-nli"]
  hatebr_offensive:
    benchmark: hatebr_offensive
    col_name: HateBR
    task_list:
    - hatebr_offensive
    metric: f1_macro
    few_shot: 25
    limit: null
    baseline: 50.0
    human_baseline: null
    expert_human_baseline: null
    description: "HateBR is the first large-scale expert annotated dataset of Brazilian Instagram comments for abusive language detection 
    on the web and social media. The HateBR was collected from Brazilian Instagram comments of politicians and manually annotated 
    by specialists. It is composed of 7,000 documents annotated with a binary classification (offensive 
    versus non-offensive comments)."
    link: https://arxiv.org/abs/2103.14972
    sources: ["https://huggingface.co/datasets/eduagarcia/portuguese_benchmark", "https://github.com/franciellevargas/HateBR", "https://huggingface.co/datasets/ruanchaves/hatebr"]
  portuguese_hate_speech:
    benchmark: portuguese_hate_speech
    col_name: PT Hate Speech
    task_list:
    - portuguese_hate_speech
    metric: f1_macro
    few_shot: 25
    limit: null
    baseline: 47.9
    human_baseline: null
    expert_human_baseline: null
    description: "Portuguese dataset for hate speech detection composed of 5,668 tweets with binary annotations (i.e. 'hate' vs. 'no-hate')"
    link: https://aclanthology.org/W19-3510/
    sources: ["https://huggingface.co/datasets/eduagarcia/portuguese_benchmark", "https://github.com/paulafortuna/Portuguese-Hate-Speech-Dataset", "https://huggingface.co/datasets/hate_speech_portuguese"]
  tweetsentbr:
    benchmark: tweetsentbr
    col_name: tweetSentBR
    task_list:
    - tweetsentbr
    metric: f1_macro
    few_shot: 25
    limit: null
    baseline: 32.8
    human_baseline: null
    expert_human_baseline: null
    description: "TweetSentBR is a corpus of Tweets in Brazilian Portuguese. 
    It was labeled by several annotators following steps stablished on the literature for 
    improving reliability on the task of Sentiment Analysis. Each Tweet was annotated 
    in one of the three following classes: Positive, Negative, Neutral."
    link: https://arxiv.org/abs/1712.08917
    sources: ["https://bitbucket.org/HBrum/tweetsentbr"]