VedMedTest/vetbot-v3

VetBot V3 - Qwen2.5-72B LoRA Adapter

Fine-tuned Qwen2.5-72B-Instruct for veterinary medicine Q&A.

Training Summary

Parameter	Value
Base Model	Qwen/Qwen2.5-72B-Instruct
Method	LoRA (r=128, alpha=256)
Trainable Params	1.68B / 74.4B (2.26%)
Hardware	4x NVIDIA H200 (141GB each)
Training Time	8.46 hours
Epochs	3
Train Samples	41,506
Final Loss	0.572
Token Accuracy	~94%

Files

• fsdp_checkpoint/ - FSDP2 distributed checkpoint
• adapter_config.json - LoRA configuration
• tokenizer* - Qwen2.5-72B tokenizer

Note

This is an FSDP2 checkpoint format. Load with HuggingFace accelerate library.