v1.1 just shipped — more than 2× the rows, same price
The only purpose-built
European Portuguese support dataset
Most "Portuguese" instruction datasets are actually Brazilian Portuguese. LusoSupport-PT — European Portuguese Customer Support Dataset (Individual Licence) is written exclusively in pt-PT — validated, structured, and ready to fine-tune.
"instruction": "Classifica a intenção principal do cliente e responde apenas com um objeto JSON."
"input": "Bom dia, preciso da fatura referente ao pagamento deste mês."
"output": {"intent": "invoice_request",
"urgency": "low",
"domain": "billing_accounts"}8
Domains
8
Task types
18
Customer intents
100%
pt-PT validated
pt-PT ≠pt-BR
Vocabulary, verb conjugation, and register genuinely diverge. Models trained on pt-BR data reliably produce phrasing that reads as wrong to a Portugal-based audience — every row here is checked against it.
Typical "Portuguese" datasets (pt-BR)
-
✕
celular -
✕
senha -
✕
nota fiscal
LusoSupport-PT (pt-PT, always)
- ✓ telemóvel
- ✓ palavra-passe
- ✓ fatura
Coverage that goes deep
Real support scenarios across 8 domains, weighted toward where support volume actually happens.
By domain
By task type
18
Customer intents
3
Difficulty levels
4
Channels
4
Export formats
Ready for your pipeline
No conversion scripts needed — pick the format that fits your stack.
JSONL
Raw instruction/input/output rows
Alpaca JSONL
Drop-in for Axolotl / LLaMA-Factory / unsloth
CSV
For spreadsheets and quick QA
Parquet
Efficient columnar loading at scale
Who this is built for
If any of these sound like you, the dataset pays for itself the first time you use it.
SaaS & e-commerce teams serving Portugal
Fine-tune a support model or copilot that replies in real pt-PT — telemóvel, palavra-passe, fatura — instead of Brazilian phrasing that reads as foreign to your customers.
Agencies & freelance ML consultants
Skip weeks of manual data collection for a client's Portuguese support-automation project — 10,828 ready-labelled rows across 8 business domains, licensed for commercial fine-tuning.
Chatbot & helpdesk copilot builders
Bootstrap intent detection, urgency triage, and next-action suggestion for a support bot without hand-labelling a single ticket.
Researchers & NLP practitioners
Benchmark pt-PT generation quality, run few-shot prompting experiments, or evaluate variant drift (pt-PT vs. pt-BR) with a clean, richly annotated reference set.
Enterprises with a Portugal support desk
Use `rewrite_professional` rows to train a tone-consistency layer so every agent's reply — regardless of who typed it — reads polished and on-brand.
Indie devs building RAG support tools
Index approved rows as a retrieval corpus so a general-purpose LLM answers in grounded, on-tone pt-PT — no fine-tuning required.
Integrate in minutes
Load it with the tools you already use — no custom parsing.
from datasets import load_dataset
ds = load_dataset("json",
data_files="lusosupport_pt_v1.jsonl", split="train")
# 10,828 rows, ready for Unsloth /
# LLaMA-Factory / Axolotl / OpenAI FT
split = ds.train_test_split(test_size=0.1)# No fine-tuning needed — steer GPT-4 / Claude / Gemini
shots = [r for r in rows
if r["task_type"] == "response_generation"
and r["domain"] == "ecommerce"][:5]
system_prompt = build_prompt(shots)
# Model now replies in real pt-PT registerFull working code for Unsloth, LLaMA-Factory, OpenAI fine-tuning, LangChain/ChromaDB RAG, and evaluation: Integration guide →
Questions
Is this really pt-PT and not pt-BR?
Yes. Every row runs through an automated validator that rejects banned pt-BR vocabulary (celular, senha, nota fiscal, etc.) and confirms the pt-PT variant tag before release.
What's the licence?
This Individual licence covers personal, research, and non-commercial internal use. A separate Commercial licence is available on Gumroad for production/commercial fine-tuning.
Can I try before buying?
Yes — a free 200-row sample covering all domains and task types is on Hugging Face under CC BY 4.0.
Do I get future updates?
Yes, all future v1.x releases (more rows, more domains, harder edge cases) are included at no extra cost. v1.1 already shipped: rows doubled from 5,162 to 10,828 with no price change for existing buyers.
LusoSupport-PT — European Portuguese Customer Support Dataset (Individual Licence)
10,828 validated rows, ready to fine-tune. Early adopter pricing won't last.
Secure checkout via Gumroad · Instant download