AMALIA-ExtraiConheCiencia-PT

Model Description

AMALIA-ExtraiConheCiencia-PT is a fine-tuned version of the AMALIA 9B European Portuguese large language model, adapted for Scientific Information Extraction (SciIE) in Portuguese.

The model was fine-tuned on SCIERC-PT, a European Portuguese translation and curation of the SCIERC dataset. It is designed to perform joint extraction of scientific entities and relations from articles abstracts, producing structured outputs suitable for Named Entity Recognition (NER), Relation Extraction (RE), and downstream Scientific Knowledge Graph construction.

Base Model

• Base model: AMALIA-9B-1225-SFT
• Language: European Portuguese (PT-PT)
• Architecture: Decoder-only generative LLM
• Fine-tuning method: QLoRA
• Fine-tuning Task: Joint Scientific Information Extraction

Intended Use

This model is intended for:

• Scientific Information Extraction in Portuguese:
  • Named Entity Recognition from scientific abstracts
  • Relation Extraction from scientific abstracts
• Structured extraction for Scientific Knowledge Graph construction

Input Format

The model expects a European Portuguese scientific abstract and task instructions specifying the entity and relation schema.

Example input:

Extrai entidades e relações do texto científico usando o esquema fornecido.

Tipos de entidade: Tarefa, Método, Métrica, Material, Outros Termos Científicos, Genérico.
Tipos de relação: Usado-para, Característica-de, Hipónimo-de, Parte-de, Avaliado-para, Comparado-com, Conjunção.

Texto:
{text}

Output Format

The model was fine-tuned to generate structured outputs containing scientific entities, entity types, and relations.

The expected output corresponds to structured triples.

Example output:

[
  {
    "sujeito": "<subject entity text>",
    "tipo_sujeito": "<subject entity type>",
    "relação": "<relation type>",
    "objeto": "<object entity text>",
    "tipo_objeto": "<object entity type>"
  }
  ...
]

Training Data

The model was fine-tuned on SCIERC-PT, a European Portuguese version of SCIERC.

SCIERC-PT contains computer science abstracts annotated with scientific entities and semantic relations.

Entity Types

English	Portuguese
Task	Tarefa
Method	Método
Evaluation Metric	Métrica de Avaliação
Material	Material
Other Scientific Term	Outros Termos Científicos
Generic	Genérico

Relation Types

English	Portuguese
Used-for	Usado-para
Feature-of	Característica-de
Hyponym-of	Hipónimo-de
Part-of	Parte-de
Compare	Comparado-com
Conjunction	Conjunção
Evaluate-for	Avaliado-para

Training Procedure

The model was adapted using QLoRA for parameter-efficient supervised fine-tuning. Fine-tuning was performed using the joint extraction formulation, where the model receives a Portuguese scientific abstract and generates structured triples containing entities, entity types, and relations.

Training Configuration

Parameter	Value
Base model	AMALIA-9B-1225-SFT
Fine-tuning method	QLoRA
Quantization (training only)	4-bit NF4
Double quantization	Yes
Compute dtype	bf16
Training precision	bf16
Maximum sequence length	2048
Training batch size	1
Gradient accumulation steps	8
Effective batch size	8
Epochs	5
Learning rate	2e-4
LoRA rank (r)	8
LoRA alpha	16
LoRA dropout	0.05
Target modules	all linear layers
Bias	none
Task type	Causal language modeling
Evaluation strategy	Per epoch
Save strategy	Per epoch
Best model selection	Lowest validation loss
Early stopping patience	2 epochs

The model was trained using the training split of SCIERC-PT and validated on the development split. The best checkpoint was selected according to validation loss.

Evaluation

The model was evaluated on the SCIERC-PT test split using F1-score under a strict exact-match evaluation protocol.

In the fine-tuned setting, AMALIA-ExtraiConheCiencia-PT achieved:

Task	Macro-F1
NER	0.44
RE	0.16

More Information

For more information about the model, the fine-tuning dataset and process, and its use in Portuguese Scientific Information Extraction experiments, see our paper:

• Paper: Evaluating Generative Large Language Models for Portuguese Scientific Information Extraction (NSLP 2026)
• https://lrec.elra.info/lrec2026-ws-nslp-15