UdayGattu23/qwen2.5-7b-finetuned-argus

Qwen2.5-7B Finetuned on Argus Dataset

This model is a finetuned version of Qwen2.5-7B using LoRA with rank 128.

Training Details

• Base Model: Qwen/Qwen2.5-7B
• Training Method: LoRA (rank=128, alpha=256)
• Dataset: 27,997 text samples
• Epochs: 2 (best checkpoint from epoch 1)
• Batch Size: 16 (effective)
• Learning Rate: 5e-5
• Hardware: A100 GPU

Training Results

• Epoch 1: Training Loss: 1.301, Validation Loss: 1.589 (best)
• Epoch 2: Training Loss: 1.699, Validation Loss: 1.826

Available Formats

• PyTorch: Original model weights
• GGUF: Multiple quantization levels available
  • Q8_0: Highest quality (7.5GB)
  • Q6_K: Very high quality (5.5GB)
  • Q5_K_M: High quality (4.8GB)
  • Q4_K_M: Good quality (3.8GB)
  • Q4_0: Acceptable quality (3.5GB)

Usage

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("UdayGattu23/qwen2.5-7b-finetuned-argus")
tokenizer = AutoTokenizer.from_pretrained("UdayGattu23/qwen2.5-7b-finetuned-argus")

With llama.cpp (GGUF)

./main -m qwen2.5-7b-finetuned-Q4_K_M.gguf -p "Your prompt here"

License

Apache 2.0