AG News RoBERTa Full-Text Seed 42 50pct
This model is a fine-tuned roberta-base sequence classifier for the AG News four-class classification task.
Model details
- Base model:
roberta-base - Dataset: AG News
- Task: four-class news classification
- Language: English
- Random seed: 42
- Training representation: full text
- Training fraction: 0.5
- Training examples: 57000
- Validation examples: 6,000
- Test examples: 7600
- Maximum sequence length: 128 RoBERTa tokens
Labels
| Label ID | Label |
|---|---|
| 0 | World |
| 1 | Sports |
| 2 | Business |
| 3 | Sci/Tech |
Evaluation on AG News test set
| Metric | Value |
|---|---|
| Test accuracy | 0.947237 |
| Test macro F1 | 0.947215 |
| Accuracy 95% Wilson CI | 0.941981–0.952041 |
| Misclassified examples | 401 |
| Single-example latency mean | 12.267 ms |
| Training time | 1285.03 s |
Per-class F1
| Class | F1 |
|---|---|
| World | 0.955263 |
| Sports | 0.988421 |
| Business | 0.919222 |
| Sci/Tech | 0.925955 |
Training setup
- Learning rate: 2e-5
- Epochs: 3
- Train batch size: 16
- Eval batch size: 32
- Weight decay: 0.01
- Warmup ratio: 0.1
- Validation split: 5% stratified split from the original AG News training set
- Training pool after validation holdout: 114,000 examples
Loading example
repo_id = "martian786/agnews-roberta-seed42-fulltext-50pct"
Use AutoTokenizer.from_pretrained(repo_id) and AutoModelForSequenceClassification.from_pretrained(repo_id).
Limitations
This model was fine-tuned on AG News and is intended for controlled academic comparison rather than production deployment.
- Downloads last month
- 2
Model tree for martian786/agnews-roberta-seed42-fulltext-50pct
Base model
FacebookAI/roberta-baseDataset used to train martian786/agnews-roberta-seed42-fulltext-50pct
Evaluation results
- Test Accuracy on AG Newstest set self-reported0.947
- Test Macro F1 on AG Newstest set self-reported0.947