roberta-babe-baseline

Best-fold checkpoint from a 5-fold RoBERTa-base reproduction of BABE sentence-level media bias classification.

• Training code: https://github.com/vulonviing/babe-roberta-baseline
• Source dataset: https://huggingface.co/datasets/mediabiasgroup/BABE
• Released checkpoint: models/fold_0/checkpoint-532
• Selected checkpoint: fold_0 with macro-F1 0.876
• Summary statement: trained on 80% of BABE, 5-fold CV mean: 0.857 +- 0.012

Model details

Item	Value
Base model	roberta-base
Task	Sentence-level media bias classification
Labels	non-biased, biased
Max sequence length	128
Epochs	4
Learning rate	2e-05
Batch size	16 train / 32 eval
Weight decay	0.01
Warmup ratio	0.1
Random seed	42

Cross-validation summary

Metric	Mean +- Std
Macro-F1	0.857 +- 0.012
Accuracy	0.858 +- 0.012
Precision (macro)	0.856 +- 0.011
Recall (macro)	0.859 +- 0.012
Biased F1	0.869 +- 0.011

Per-fold macro-F1 values in the repo: 0.876, 0.854, 0.845, 0.852, 0.856.

Held-out quick-run reference

Metric	Score
Macro-F1	0.870
Accuracy	0.872
Precision (macro)	0.870
Recall (macro)	0.872
Biased F1	0.884

Confusion matrix from the held-out quick run (n=468):

	Pred non-biased	Pred biased
True non-biased (207)	180	27
True biased (261)	33	228

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer

repo_id = 'vulonviing/roberta-babe-baseline'
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)