RL4LMs
RL4LMs copied to clipboard
Question about the classifier used for IntentAccuracyDailyDialog.
According to the source code of class IntentAccuracyDailyDialog(BaseMetric), the intent likelihood of utterances on DailyDialog is computed by rajkumarrrk/roberta-daily-dialog-intent-classifier
.
However, according to the config.json
of this classifier, it is used for emotion classification, with four labels: joy, optimism, anger, and sadness, while the intent labels on DailyDialog should be Inform, Questions, Directives, and Commissive instead.
So my question is: Is this classifier already fine-tuned on intent classification of DailyDialog utterances?
Empirically, i obeserve that the classification results of ground truth utterances in DailyDialog by this classifier are unbalanced and not well-aligned to the labelled intent distribution, as shown below.
- classification results on test set
label-0 | label-1 | label-2 | label-3 | Intent Accuracy | |
---|---|---|---|---|---|
classification on ground truth | 0.7102 | 0.0055 | 0.0275 | 0.2071 | 0.6147 |
intent labels in DailyDialog | 0.4988 | 0.2231 | 0.1565 | 0.1213 | - |
classification on SFT generation | 0.5363 | 0.1591 | 0.0944 | 0.2100 | 0.4034 |