Primera-Summarization-Council-PT is an abstractive text summarization model based on primera, fine-tuned to produce concise and informative summaries of discussion subjects from Portuguese municipal meeting minutes.
The model was trained on a curated and annotated corpus of official municipal meeting minutes covering a variety of administrative and political topics at the municipal level.

Try out the model: Hugging Face Space Demo

Key Features

🧾 Abstractive Summarization – Generates natural, human-like summaries rather than extracts.
🇵🇹 European Portuguese – Optimized for official and administrative Portuguese.
🏛️ Domain-Specific – Trained on municipal meeting minutes and administrative discussions.
⚙️ Fine-tuned primera – Built upon allenai/primera using supervised fine-tuning.
🧠 Fact-Aware Generation – Produces short summaries that preserve factual content.

Model Details

Architecture: allenai/PRIMERA
Base Model: Longformer Encoder-Decoder (extension of BART)
Task: Abstractive summarization (text → summary)
Framework: Hugging Face Transformers (PyTorch)
Tokenizer: Longformer/BART tokenizer (English vocabulary reused for Portuguese text)
Max Input Length: 4096 tokens
Max Summary Length: 128 tokens
Training Objective: Conditional generation (cross-entropy loss)
Dataset: Portuguese municipal meeting minutes annotated with summaries

How It Works

The model receives a discussion subject of a municipal meeting and outputs a short, coherent summary highlighting:

The main subject or topic of discussion
Any decisions, motions, or proposals made
The entities or departments involved

Example Usage

python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_name = "anonymous12321/Primera-Summarization-Council-PT"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

text = """
17. PROCESSO DE OBRAS N.º ***** -- EDIFIC\nPelo Senhor Presidente foi presente a esta reunião a informação n.º ****** da Secção de Urbanismo e Fiscalização -- Serviço de Obras Particulares que se anexa à presente ata. \nPonderado e analisado o assunto o Executivo Municipal deliberou por unanimidade aprovar as especialidades relativas ao processo de obras n.º ***** -- EDIFIC.
"""

inputs = tokenizer(text, return_tensors="pt", max_length=1024, truncation=True)
summary_ids = model.generate(**inputs, max_length=128, num_beams=4, early_stopping=True)
print(tokenizer.decode(summary_ids[0], skip_special_tokens=True))

🧾 Model Output

Output:

"O Executivo Municipal aprovou, por unanimidade, as especialidades relativas a um processo de obras particulares."

📊 Evaluation Results

Quantitative Metrics (on held-out test set)

Metric	Score	Description
ROUGE-1	0.632	Unigram overlap between generated and reference summaries
ROUGE-2	0.500	Bigram overlap
ROUGE-L	0.577	Longest common subsequence overlap
BERTScore (F1)	0.846	Semantic similarity between summary and reference

⚙️ Training Details

Pretrained Model: allenai/primera
Optimizer: AdamW (default in Hugging Face Trainer)
Learning Rate: 2e-5
Batch Size: 4
Epochs: 3
Scheduler: Linear warmup
Loss Function: Cross-entropy
Evaluation Metrics: ROUGE (computed on validation set every 100 steps)
Evaluation Strategy: Step-based evaluation (eval_steps=100)
Weight Decay: 0.01
Mixed Precision (fp16): Enabled when CUDA is available
Chunking: Implemented with max_length=512 and stride=256 for hierarchical input segmentation
Target (summary) Max Length: 128 tokens

📚 Dataset Description

The model was trained on a specialized dataset of Portuguese municipal meeting minutes, consisting of:

Discussion Subjects from official municipal meeting minutes.
Decisions and deliberations across departments (urban planning, finance, education, etc.)
Expert-annotated summaries per discussion segment

Dataset sources include:

Six Portuguese municipalities meeting minutes

⚠️ Limitations

Language Restriction: The model is optimized for Portuguese; performance may degrade in other languages.
Domain Dependence: Best suited for administrative and institutional texts; less effective on informal or creative writing.
Length Sensitivity: Very long transcripts (>1024 tokens) are truncated; chunking may be needed for full documents.
Generalization: While robust within-domain, it may underperform on unseen domains or vocabulary.

📄 License

This model is released under the
Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0).