CHAST

This model is a fine-tuned version of lmsys/vicuna-13b-v1.5-16k. For more details, please refer to the paper: https://arxiv.org/pdf/2405.05378

Model description

Computes Covert Harms and Social Threats (CHAST) metrics for conversational data. For more details, please refer to the paper: https://arxiv.org/pdf/2405.05378

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 8
• eval_batch_size: 64
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• num_epochs: 2.0
• mixed_precision_training: Native AMP

Framework versions

• PEFT 0.9.0
• Transformers 4.37.2
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2

Reference and Citation

Dammu, P. P. S., Jung, H., Singh, A., Choudhury, M., & Mitra, T. (2024). "They are uncultured": Unveiling covert harms and social threats in LLM generated conversations. arXiv. https://arxiv.org/abs/2405.05378.