duarteocarmo/fineweb2-bagaco2

Bagaço2 🍷🇵🇹

A pretraining dataset for European Portuguese.

Methodology:

• Takes Portuguese split of CulturaX
• Uses this classifier to only keep PT-PT docs
• Adds educational score + content category to each row, by running two additional classifiers

33M documents in 460 parquet shards (~ 37 GB)

Classifiers

Statistics & Counts

• CulturaX filter using datatrove:

"stats": {
    "total": 199737979,
    "dropped": 166679183,
    "forwarded": 33058796,
    "doc_len": {
        "total": 90594584215,
        "n": 33058796,
        "mean": 2740.4078543876713,
        "variance": 36957836.43497764,
        "std_dev": 6079.295718664921,
        "min": 209,
        "max": 668957
    }

E.g., 33M docs passed the PT-PT test.

(below are approximate, HF was ratelimiting me)

Educational score	Count	Percentage
0	~14.5M	45.20%
1	~11.0M	34.34%
2	~4.4M	13.70%
3	~1.9M	6.04%
4	~0.2M	0.72%

Category	Count	Percentage
Lifestyle	~5.9M	18.46%
News	~5.0M	15.47%
Business	~4.7M	14.77%
Society	~4.0M	12.51%
Arts	~3.8M	11.84%
Sports	~3.0M	9.33%
Health	~2.4M	7.52%
Games	~2.0M	6.27%
Science	~1.2M	3.83%

Portuguese variety filter

Filtering was done with duarteocarmo/fasttext-euptvid, a fastText classifier for Portuguese variety identification.

• Task: PT-PT vs PT-BR classification
• Kept label: __label__PT_PT
• Threshold: 0.7
• Model file: model_quantized.ftz (quantized version)

See model card.

Educational score classifier

Each document is assigned an educational quality score.

Reference data

• Labeled samples: 30,000
• Labeling model: Qwen3 235B A22B
• Reference file: classification/bagaco_reference_educational_score_0to5_qwen3_235b_30000.parquet

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)

Validation report

Score	Precision	Recall	F1-Score	Support
0	0.66	0.82	0.73	2,056
1	0.83	0.54	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]
 [ 846 1840  564  172   16]
 [   7   65  189  101    7]
 [   1    3   40   78    9]
 [   0    0    3    3    0]]

Category classifier

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

Reference data

• Labeled samples: 3,500
• Labeling model: Gemini 2.5 Flash Lite
• Reference file: classification/bagaco_reference_category_9class_gemini25flashlite_3500.parquet

Classifier

• Embeddings: intfloat/multilingual-e5-small
• Model: Logistic Regression (C=1.0, class_weight='balanced')
• Validation: 20% held-out split (train=2,800, test=700)

Validation report

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