persuasive_essays_distilbert_uncased

Model description

This model is a fine-tuned version of distilbert-base-uncased on the emnlp2017-claim-identification/persuasive_essays dataset. It achieves the following results on the evaluation set:

• Loss: 0.4314
• Accuracy: 0.8001
• Macro F1: 0.7510
• Claim F1: 0.6403

Intended uses & limitations

Text classification for claims on full sentences. The model perfoms better at in-domain classification. Cross-domain classification is severely limited.

Training and evaluation data

Based on Stab and Gurevych (2017) persuasive essays corpus, preprocessed by Daxenberger et al. (2017).

Original dataset

• docs: 402
• tokens: 147,271
• total instances: 7,116 (65 duplicates)
  • #claims: 2,108 (29.62%)

Trimmed datast used for training

• total instances: 7051 (65 duplicates removed)
  • #claims: 2093 (29.68%)
• train/test split: 80/20, stratified

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 2
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	Macro F1	Claim F1
No log	1.0	353	0.4486	0.7810	0.7447	0.6485
0.4521	2.0	706	0.4314	0.8001	0.7510	0.6403

Framework versions

• Transformers 4.37.2
• Pytorch 2.2.0
• Datasets 2.17.0
• Tokenizers 0.15.2