Text samples from the AG News benchmark across news categories curated for error analysis and label correction workflows. It facilitates the identification of mislabeled instances by comparing original labels against model loss metrics as described in the Rubrix tutorial.
Use Cases
- Identify potential mislabels by calculating the loss between model predictions and the provided labels.
- Benchmark label cleaning tools by comparing the original labels with the human-verified ag_news_corrected_labels dataset.
- Improve model performance by replacing noisy labels with corrected values from the associated tutorial dataset.
Strengths
- Derived from the AG News dataset for text classification.
- Integrated with a tutorial for loss-based error analysis in the Rubrix/Argilla ecosystem.
- Paired with the ag_news_corrected_labels dataset for ground-truth verification.
- Includes a link to a specific tutorial on error analysis using loss metrics.