duarteocarmo/fineweb2-bagaco

Bagaço 🍷🇵🇹

Bagaço is a pretraining dataset for European Portuguese. It filters the Fineweb2 dataset to URLs from Portuguese domains (e.g., .pt/). Each document is classified into one of 9 categories and scored for educational quality.

Filtering

• Source: HuggingFaceFW/fineweb-2, subset por_Latn, split train
• Filter: URLs containing .pt/ (Portuguese top-level domain)

Document classification

Each document is classified into one of 9 categories: Society, Arts, Business, Science, Sports, Lifestyle, Health, Games, News.

Labeling:

• Model: Gemini 2.5 Flash Lite
• Labeled samples: 3,500
• Prompt:

System: Classify the Portuguese web text into one category.
User: <first 800 chars of document>
Response format: { category: Society | Arts | Business | Science | Sports | Lifestyle | Health | Games | News }

Classifier:

• Embeddings: intfloat/multilingual-e5-small
• Model: Logistic Regression (C=1.0, class_weight='balanced')
• Validation: 20% held-out split (700 samples)

Category	Precision	Recall	F1-Score	Support
Arts	0.60	0.83	0.70	59
Business	0.81	0.78	0.79	131
Games	0.75	0.91	0.82	23
Health	0.77	0.87	0.81	53
Lifestyle	0.81	0.75	0.78	111
News	0.80	0.71	0.75	131
Science	0.42	0.87	0.57	15
Society	0.72	0.57	0.64	101
Sports	0.89	0.87	0.88	76
Accuracy			0.76	700
Macro Avg	0.73	0.80	0.75	700
Weighted Avg	0.77	0.76	0.76	700

Educational score

Each document is assigned an educational quality score (0-5).

Labeling:

• Model: Qwen3 235B A22B
• Labeled samples: 30,000
• Prompt (adapted from FineWeb-Edu):

Below is an extract from a web page. Evaluate whether the page has a high educational value
and could be useful in an educational setting for teaching from primary school to grade school
levels using the additive 5-point scoring system described below. The text will be in Portuguese.
Evaluate its educational value based on content quality, not language.

- 1 point: basic information relevant to educational topics, even with ads/promotional material.
- 2 points: addresses elements pertinent to education but doesn't align closely with standards.
- 3 points: appropriate for educational use, introduces key concepts relevant to school curricula.
- 4 points: highly relevant for grade school education, clear writing, substantial content.
- 5 points: outstanding educational value, perfectly suited for primary/grade school teaching.

The extract: <first 1500 chars of document>

After examining the extract, briefly justify your total score (up to 100 words)
and provide the educational score (0-5).
Response format: { justification: str, educational_score: int }

Classifier:

• Embeddings: intfloat/multilingual-e5-small
• Model: Logistic Regression (C=1.0, class_weight='balanced')
• Validation: 20% held-out split (train=24,000, test=6,000)

Score	Precision	Recall	F1-Score	Support
0	0.66	0.82	0.73	2,056
1	0.83	0.53	0.65	3,438
2	0.23	0.51	0.31	369
3	0.21	0.60	0.31	131
4	0.00	0.00	0.00	6
Accuracy			0.63	6,000
Macro Avg	0.39	0.49	0.40	6,000
Weighted Avg	0.72	0.63	0.65	6,000

Confusion Matrix:
[[1679  319   40   15    3]
 [ 847 1839  564  172   16]
 [   7   65  189  101    7]
 [   1    3   40   78    9]
 [   0    0    3    3    0]]

Notes

• references contains the labeled datasets used to train the classifiers
• scripts contains the scripts used to train classifiers, classify documents, and process the dataset