MentaLLaMA provides open-source instruction-following large language models for interpretable mental health analysis on social media. It targets researchers and developers needing to analyze mental health discourse and generate explanations, offering a novel dataset and benchmark for this specialized domain.

How It Works

MentaLLaMA is built upon LLaMA and Vicuna foundation models, fine-tuned on the Interpretable Mental Health Instruction (IMHI) dataset. This dataset comprises 105K instruction samples across 8 mental health analysis tasks derived from public social media data. The models are designed to follow instructions for mental health analysis and provide high-quality, interpretable explanations for their predictions.

Quick Start & Requirements

• Installation: Use Hugging Face Transformers library.
• Dependencies: Python, PyTorch, Transformers, PEFT (for LoRA models). GPU recommended for inference.
• Model Loading (Example):

  from transformers import LlamaTokenizer, LlamaForCausalLM
  tokenizer = LlamaTokenizer.from_pretrained(MODEL_PATH)
  model = LlamaForCausalLM.from_pretrained(MODEL_PATH, device_map='auto')
  

• MentaLLaMA-33B-lora: Requires downloading Vicuna-33B separately and placing it in ./vicuna-33B.
• Resources: Larger models (13B, 33B) require significant VRAM.
• Docs: MentaLLaMA Paper, Evaluation Paper

Highlighted Details

• Offers 5 model checkpoints: MentaLLaMA-33B-lora, MentaLLaMA-chat-13B, MentaLLaMA-chat-7B, MentalBART, MentalT5.
• Includes the IMHI dataset (105K instruction samples) and a benchmark (19K test samples) for interpretable mental health analysis.
• Provides 10 pre-trained classifiers (based on MentalBERT) for evaluating model output correctness on the IMHI benchmark.
• Supports evaluation using BART-score for explanation quality.

Maintenance & Community

• Active development with recent updates in March 2024 (test data release) and February 2024 (paper acceptance).
• Primary contributors are affiliated with the National Centre for Text Mining and The University of Manchester.

Licensing & Compatibility

• Licensed under the MIT License.
• Permits commercial use and linking with closed-source projects.

Limitations & Caveats

• The project is strictly for non-clinical research; it does not provide diagnosis or advice.
• Users assume all risk; authors disclaim responsibility for errors or consequences.
• LLMs may introduce biases, incorrect predictions, or inappropriate explanations, posing challenges for real-world deployment.