amalia-llm/EGSciQA-ptPT-V1

EGSciQA-ptPT-V1

EGSciQA-ptPT-V1 is the first European Portuguese (pt-PT) supervised fine-tuning dataset for evidence-grounded scientific question answering and reasoning. Each example contains a system prompt, an instruction with scientific evidence, and a structured target response using <raciocinio> and <resposta> sections.

Samples were generated directly from real pt-PT scientific manuscripts available in the amalia-llm/CorEGe-PT corpus.

This repository contains the final dataset artifacts, including instructions and system prompts. It is not the raw pipeline export. Each eligible source example contributes exactly one row: the original row is not included alongside additional copies.

Splits

lite-eval is a distribution-matched 300-example subset of test; it is not counted again in the 17,921 unique examples.

Fields

The main fields are:

• —question_id, doc_id: source identifiers.
• —source_type: single_document or document_cluster.
• —question_family: question-generation family.
• —groundability: groundable, partially_groundable, or non_groundable.
• —system: system message used for SFT.
• —instruction: question and evidence in one augmented presentation.
• —response: supervised assistant target.
• —selected_answer: selected final answer before chat formatting.
• —cot_sample_index: selected candidate index in the source pipeline.
• —split, split_group: leakage controlled split metadata.
• —cluster_id: present for document-cluster examples.

Intended use

Use train for SFT or RL-Based alignment, validation for model selection and training diagnostics, lite-eval for faster development evaluations, and test only for final held-out evaluation. The lite split approximates the full test distribution over evidence type, domain, groundability, and agreement band. The model should be trained to answer exclusively from the supplied evidence, cite the evidence identifiers it uses, and explicitly state when the evidence is insufficient.

Limitations

The examples are synthetically generated from real scientific documental evidence and selected by an automated pipeline. They may retain OCR artifacts, source document errors, or imperfect generated reasoning despite automated validation. The dataset should not be treated as a substitute for expert review.