qwen2.5-7b-therapist-v2

Qwen2.5-7B-Instruct Therapist v2

Fine-tuned version of Qwen/Qwen2.5-7B-Instruct for therapeutic conversations.

Model Details

• Base Model: Qwen/Qwen2.5-7B-Instruct
• Fine-tuning Method: LoRA (Rank 64) + merged into full model
• Training Dataset: Jyz1331/therapist_conversations (250 conversations)
• Training Infrastructure: Modal A100 GPU
• Training Time: ~10 min on A100

Training Pipeline

Phase 1: SFT (LoRA r=64)

• Learning rate: 2e-4, cosine schedule, 10% warmup
• Effective batch size: 32 (4×8 gradient accumulation)
• Epochs: 3
• bf16 training, gradient checkpointing
• Trainable params: ~340M (4.5% of total)

Phase 2: Merged

• LoRA adapter merged into base model for single-file deployment

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ArkMaster123/qwen2.5-7b-therapist-v2")
tokenizer = AutoTokenizer.from_pretrained("ArkMaster123/qwen2.5-7b-therapist-v2")

messages = [{"role": "user", "content": "I'm feeling anxious about work"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Safety Limitations

This model is not a replacement for professional mental health services. For crises, please contact emergency services or a crisis helpline.

Citation

@misc{therapist-v2,
  author = {ArkMaster123},
  title = {Qwen2.5-7B-Instruct Therapist v2},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ArkMaster123/qwen2.5-7b-therapist-v2}}
}