thesa

Thesa: A Therapy Chatbot 👩🏻‍⚕️

Thesa is an experimental project of a therapy chatbot trained on mental health data and fine-tuned with the Zephyr GPTQ model that uses quantization to decrease high computatinal and storage costs.

Model description

• Model type: fine-tuned from TheBloke/zephyr-7B-alpha-GPTQ on various mental health datasets
• Language(s): English
• License: MIT

Intended uses & limitations

This model is purely experimental and should not be used as substitute for a mental health professional.

Training evaluation

Training loss: <img src="imgs/loss_27.2.24.png" alt="loss" width="550"/>

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• warmup_ratio: 0.1
• train_batch_size: 8
• eval_batch_size: 8
• gradient_accumulation_steps: 1
• seed: 35
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 10
• mixed_precision_training: Native AMP
• fp16: True

Learning rate overtime (warm up ratio was used during training): <img src="imgs/lr_27.2.24.png" alt="lr" width="550"/>

Framework versions

• Transformers 4.35.2
• Pytorch 2.1.0+cu121
• Datasets 2.16.1
• Tokenizers 0.15.1
• Accelerate 0.27.2
• PEFT 0.8.2
• Auto-GPTQ 0.6.0
• TRL 0.7.11
• Optimum 1.17.1
• Bitsandbytes 0.42.0