🤗 Hugging Face · LLM Fine-Tuning Project

Qwen2.5 QLoRA

Qwen2.5-1.5B-Instruct fine-tuned with QLoRA on the Databricks Dolly 15K instruction dataset.

🧠 Model & Training

Qwen2.5-1.5B-Instruct QLoRA LoRA r=16 Alpha=32 Dolly 15K 1 Epoch 4-bit NF4

💬 Prompt Playground

MODEL RESPONSE
Enter a prompt to test the demo interface.

📊 Held-Out Evaluation

Base Model 18.30 / 20 Average score across 20 questions
Fine-Tuned Model 17.95 / 20 Average score across 20 questions

The held-out evaluation was scored using predefined criteria: correctness, instruction following, relevance, and clarity. On this particular evaluation set, the base model scored slightly higher overall. This highlights why training loss alone is not sufficient to demonstrate improvement.

🔬 Project Details

PyTorch Transformers PEFT TRL BitsAndBytes QLoRA

The project compares the original Qwen2.5-1.5B-Instruct model against a LoRA adapter trained on instruction-following data. Evaluation includes unseen questions to examine generalization.

🔗 Project Resources

💻 GitHub Repository 🤗 Hugging Face Model