AMALIA-RacioCiencia-PT

Model Card for AMALIA-RacioCiência-PT

Model description

This model is a European Portuguese scientific question-answering assistant fine-tuned from AMALIA, more specifically, amalia-llm/AMALIA-9B-1225-SFT . It answers questions using scientific evidence supplied in the prompt and cites the relevant evidence identifiers.

The model is primarily intended for European Portuguese (pt-PT) and its training data covered three specific scientific domains:

• Engineering and Technology (EngTec)
• Exact and Natural Sciences (ExaNat)
• Medical and Health Sciences (MedSau)

Intended use

Use the model to explain, compare, or synthesize information from scientific passages that you provide. Every factual answer should be grounded in those excerpts using citations such as [E1] and [E2].

The model does not retrieve or verify documents by itself, nor was it directly fine-tuned for tool usage. A retrieval system or the user must supply the evidence.

Prompt example

System:
És um especialista em raciocínio científico em português europeu. Responde apenas com base na evidência fornecida e cita os identificadores relevantes.

User:
Pergunta:
<question>

Evidência Científica:
[E1]
Documento: <document identifier>
Secção: <section title>
Texto:
<scientific excerpt>

[E2]
Documento: <document identifier>
Secção: <section title>
Texto:
<scientific excerpt>

Expected output:

<raciocinio>
Reasoning grounded in the supplied evidence, with citations such as [E1].
</raciocinio>

<resposta>
Final evidence-grounded answer, with citations such as [E1] and [E2].
</resposta>

If the evidence is insufficient, the model should state that explicitly rather than fill the gap with external knowledge.

Training data

The model is fine-tuned on the training split of amalia-llm/EGSciQA-ptPT-V1. The training set contains 12,571 examples.

Limitations

• The model can still produce unsupported statements or incorrect citations.
• Citations identify supplied evidence blocks; they are not independently verified bibliographic references.
• Performance is strongest in the represented domains and in European Portuguese.
• The reasoning traces originate from model-generated supervision and may contain imperfections.
• The model should not be used as the sole basis for medical, safety-critical, legal, or other high-stakes decisions.

Users should inspect the cited excerpts and validate important conclusions against the original documents.