ro-political-leaning-v2

ro-political-leaning-v2

Fine-tuned dumitrescustefan/bert-base-romanian-cased-v1 for Romanian political leaning scoring (v2).

Unlike v1 (manuelraileanu/ro-political-leaning, 6-class classification on XLM-RoBERTa), this model uses a single regression head that directly outputs a leaning score on the -1.0 .. +1.0 axis (v2 labeling pipeline, gemma2:9b labels, extended training set). All 7 political classes are present (far_left is no longer merged into left).

Output

A single regression value in [-1.0, +1.0], the midpoint of the predicted leaning class:

Training

• Base model: dumitrescustefan/bert-base-romanian-cased-v1
• Labels produced with gemma2:9b (local, via Ollama), rows without a clear stance (unclear) excluded
• Dataset 2 (extended): petrematei/ro-political + readerbench/news-ro-offense
• Labeled rows: 72400 (train), 8045 (test, 10% stratified split, seed 42)
• fp16, batch_size=16, grad_accum=2, max_length=512, weight_decay=0.01
• learning_rate=6e-05, 7 epochs, MSE loss, early stopping on mae (patience=2)

Dataset sources

Results (test set)

• Eval MAE: 0.181 (on the -1..+1 axis)
• Eval RMSE: 0.303
• Exact class accuracy: 0.610
• Adjacent-class accuracy: 0.884
• Pearson correlation: 0.606

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model_id = "manuelraileanu/ro-political-leaning-v2"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id).eval()

inputs = tok("Guvernul a anunțat ieri noi măsuri de stimulare economică.", return_tensors="pt")
with torch.no_grad():
    score = model(**inputs).logits.reshape(-1)[0].item()  # -1.0 (far left) .. +1.0 (far right)