md-templates: PT-BR legal petition templates

4,941 Brazilian Portuguese petition models (initial petitions, habeas corpus, appeals, criminal briefs, …) scraped from a commercial petition-model site, split into sections and templatized by an LLM: case-specific facts are replaced with {PLACEHOLDER} variables, and each section carries the variables found, concerns, and challenges. A second config holds 919 outputs of the beautifier.py clean-up pass that failed its similarity check. Built in April 2025 for a petition-drafting assistant.

Related repos

Repo	Role
arthrod/beautify (private)	Full pipeline with every intermediate stage; rough_templates here equals its chunks_three_processed stage format
arthrod/docx-corpus	Legal-documents family: DOCX corpus
arthrod/contract_parse	Legal-documents family: contract structure parsing
arthrod/sec-ex10-exhibits	Legal-documents family: US material contracts

Configs and splits

config	split	rows	source file	description
rough_templates (default)	train	4,941	rough_templates_two.jsonl	Templatized petitions (4,855 distinct _id; 86 repeated)
rough_templates_sample	train	100	rough_templates_sample.jsonl	The first 100 rows of rough_templates, identical
failed_templates	train	919	failed_templates.jsonl	beautifier.py outputs rejected by the similarity check; all 919 _ids also appear in rough_templates

The raw JSONL files are kept. In the raw files each section is a separate key (secao_1 … secao_N, secao_N_templatized_document, …); in parquet they are ordered lists (secoes[i] = secao_{i+1}). Nested objects are JSON strings. beautifier_checkpoint.json (list of processed _ids) and the scripts beautifier.py / main.py are not data.

Schema

rough_templates (and rough_templates_sample)

column	type	description
_id	string	Document UUID (shared with the private beautify pipeline)
timestamp	string	When the page was processed (ISO 8601, Apr 2025)
full_unprocessed	string	Scraped page as Markdown, including site chrome
source_file	string	Scraped page file name (site name removed)
nome_do_documento	string	LLM-assigned document name
tipo_do_documento	string	LLM-assigned document type (e.g. Petição Inicial, Habeas Corpus)
palavras_chave	list	LLM-assigned keywords
which_instancia	string	Court / instance the model petition targets
which_party	string	Party the petition is written for (autor, réu, …)
is_document_real_template_or_can_be_good_template	bool	LLM verdict: usable as a template (True in all 4,941 rows)
full_template	string	Concatenated templatized petition with {PLACEHOLDER} variables
secoes	list	Original sections (secao_1…N) in order
secoes_templatized_document	list	Per-section templatized text
secoes_variables_identified	list (JSON items)	Per-section JSON map placeholder → description
secoes_models_concerns_on_using_as_template	list	Per-section LLM note on template suitability
secoes_templatization_challenges	list (JSON items)	Per-section JSON list of problems found

failed_templates

column	type	description
_id	string	Document UUID (shared with the private beautify pipeline)
source_file	string	Scraped page file name (site name removed)
original_type	string	Document type before beautification
beautified_title	string	Title produced by beautifier.py
beautified_type	string	Type produced by beautifier.py
beautified_markdown	string	Beautified Markdown template
sections	string (JSON)	JSON list of {heading, content}
variables_found	list	Placeholders found in the output
summary	string	One-paragraph summary
validation	string (JSON)	JSON {similarity, passed}: similarity to the source below the threshold → rejected

Examples

Strings truncated to 160 characters and lists to 3 items (…).

rough_templates

{
  "_id": "1aa1f94e-3607-4b23-9f23-471cbb3e0d20",
  "timestamp": "2025-04-27T03:23:59.113814",
  "full_unprocessed": "NaN%\n\nReceba 10% de desconto nas compras com pagamento via \n\nAvalie-nos e receba de brinde diversas petições!\n\n- _star\\_rate_\n- _star\\_rate_\n- _star\\_rate_\n- _s…",
  "source_file": "www___com_br_modelo_alegacoes_finais_falta_provas_homicidio.md",
  "nome_do_documento": "Memoriais Criminais",
  "tipo_do_documento": "Memoriais Criminais",
  "palavras_chave": [
    "Memoriais Criminais",
    "Tribunal do Júri",
    "Homicídio Doloso",
    "…"
  ],
  "which_instancia": "Vara Criminal",
  "which_party": "Réu",
  "is_document_real_template_or_can_be_good_template": true,
  "full_template": "EXCELENTÍSSIMO SENHOR DOUTOR JUIZ DE DIREITO DA **{NUMERO_VARA}ª VARA CRIMINAL** DE {CIDADE_ESTADO}\n\n**Ação Penal – Rito Especial (Tribunal do Júri)**\n\nProc. nº…",
  "secoes": [
    "NaN%\n\nReceba 10% de desconto nas compras com pagamento via \n\nAvalie-nos e receba de brinde diversas petições!\n\n- _star\\_rate_\n- _star\\_rate_\n- _star\\_rate_\n- _s…",
    "## **ALEGAÇÕES FINAIS**\n\n**“MEMORIAIS SUBSTITUTIVOS”**\n\nquanto à pretensão condenatória ostentada em desfavor de **FRANCISCO DAS QUANTAS**, já qualificado na ex…",
    "### **3  – PRELIMINARMENTE**\n\n",
    "…"
  ],
  "secoes_templatized_document": [
    "EXCELENTÍSSIMO SENHOR DOUTOR JUIZ DE DIREITO DA **{NUMERO_VARA}ª VARA CRIMINAL** DE {CIDADE_ESTADO}\n\n**Ação Penal – Rito Especial (Tribunal do Júri)**\n\nProc. nº…",
    "## **ALEGAÇÕES FINAIS**\n\n**“MEMORIAIS SUBSTITUTIVOS”**\n\nquanto à pretensão condenatória ostentada em desfavor de {NOME_PARTE_ACUSADA}, já qualificado na exordia…",
    "### **3 – PRELIMINARMENTE**",
    "…"
  ],
  "secoes_variables_identified": [
    "{\"NUMERO_DO_PROCESSO\": \"Número do processo judicial (ex: 7777.33.2222.5.06.4444).\", \"NUMERO_VARA\": \"Número da vara criminal onde o processo tramita (ex: 00ª).\",…",
    "{\"HORARIO_DO_CRIME\": \"Horário em que o crime ocorreu.\", \"DATA_RECEBIMENTO_PECA\": \"Data em que a peça acusatória foi recebida.\", \"DATA_DO_CRIME\": \"Data em que o …",
    "{}",
    "…"
  ],
  "secoes_models_concerns_on_using_as_template": [
    "O modelo é bom, mas é importante verificar a adequação da jurisprudência e doutrina utilizada para garantir que estejam atualizadas e alinhadas com as últimas d…",
    "O documento é adequado para uso como modelo, pois a estrutura e a terminologia jurídica são preservadas, e apenas informações específicas do caso são substituíd…",
    "This section is a standard heading and does not contain any specific information that needs to be templatized. Therefore, it is suitable for use as is.",
    "…"
  ],
  "secoes_templatization_challenges": [
    "[]",
    "[]",
    "[]",
    "…"
  ]
}

failed_templates

{
  "_id": "2f3478e9-e375-430c-bb36-cce7cbb719b9",
  "source_file": "www___com_br_modelos_de_peticoes_autorizacao_para_citacao_apos_as_20_horas_e_em_domingos_e_feriados_novo_cpc.md",
  "original_type": "Petição",
  "beautified_title": "Requerimento de Autorização para Citação Fora do Horário Legal",
  "beautified_type": "Petição/Requerimento Intermediário",
  "beautified_markdown": "# Requerimento de Autorização para Citação Fora do Horário Legal\n\n_Petição incidental requerendo autorização judicial para que a citação do Requerido seja reali…",
  "sections": "[{\"heading\": \"Endereçamento\", \"content\": \"EXCELENTÍSSIMO(A) SENHOR(A) DOUTOR(A) JUIZ(A) DE DIREITO DA {NUMERO_DA_VARA}ª VARA CÍVEL DA COMARCA DE {NOME_DA_COMARC…",
  "variables_found": [
    "{NUMERO_DA_VARA}",
    "{NOME_DA_COMARCA}",
    "{NOME_COMPLETO_PARTE_AUTORA}",
    "…"
  ],
  "summary": "Petição incidental requerendo autorização judicial para que a citação do Requerido seja realizada fora do horário forense normal (após as 20h ou em domingos/fer…",
  "validation": "{\"similarity\": 46.17363344051447, \"passed\": false}"
}

Usage

from datasets import load_dataset

templates = load_dataset("arthrod/md-templates", "rough_templates", split="train")
print(templates[0]["full_template"][:500])

Limitations

• Scraped commercial content, no license. The source pages come from a paid petition-model site; redistribution rights were never granted. Treat as research-only.
• Contains real personal data (not removed). The scraped pages include the store's customer-review sections and the authors' bylines, so rough_templates has, besides obvious placeholders (fulano@…, OAB/PP 112233): about 128 distinct real-looking e-mail addresses (many truncated or masked by the site) in 135 rows, about 71 distinct OAB registration numbers from real states next to lawyers' names in 107 rows, and 76 CPF-formatted numbers, 4 of which pass the CPF checksum (≈0.8 expected by chance). failed_templates (the LLM output) has no e-mails or OAB numbers. Do not redistribute or train on rough_templates without scrubbing these.
• Site noise. full_unprocessed and the first sections contain page chrome (discount banners, star ratings). Use full_template / secoes_templatized_document for clean text.
• LLM labels are unchecked. The template verdict is True for every row, including 184 rows typed "Anúncio Publicitário" (ads), so it does not filter anything.
• palavras_chave was stored as a Python-list string in some raw rows and was parsed into a list here.
• 86 _ids repeat in rough_templates; deduplicate by _id before training.

License

No license was declared by the source site or the uploader (license: other).