Datasets:
Welcome + why LusoSupport-PT exists (pt-PT vs pt-BR)
👋 Welcome to the LusoSupport-PT community discussion!
Why this dataset exists: most "Portuguese" instruction/fine-tuning datasets available are actually pt-BR (Brazilian Portuguese), or an inconsistent mix of both. Vocabulary, verb conjugation (tu vs você), and formal register genuinely diverge between pt-PT and pt-BR — models fine-tuned on pt-BR data reliably produce phrasing that reads as wrong to a Portugal-based audience.
LusoSupport-PT is a customer-support instruction dataset written exclusively in European Portuguese (pt-PT), covering 8 domains (e-commerce, travel, subscriptions, telecom, SaaS, utilities, billing/accounts, marketplace) and 8 task types (intent classification, urgency classification, FAQ answering, email replies, summarization, response generation, professional rewriting, next-action suggestion).
Every row passes an automated pt-PT validator that rejects banned pt-BR vocabulary ("celular", "senha", "nota fiscal", etc.) and enforces schema/structure. This Lite sample (200 rows) is free and CC BY 4.0 — the full 5,162-row dataset is available under individual/commercial licence (link in the dataset card).
Feedback welcome: if you're working on Portuguese NLP, chatbots, or fine-tuning for a PT-PT audience, I'd love to hear your use case, and any gaps you'd want covered in a v1.1 (more domains, edge cases, multi-turn dialogues are already on the roadmap).
GitHub (schema, scripts, docs): https://github.com/ariazevedopt/pt-ai-instruction-dataset