Python code generation — LLaMA 3.2 3B + QLoRA

A 9.2M-parameter LoRA adapter (0.285% of the model), trained in 1.88 hours on a free T4.

HumanEval pass@1 40.5% → 54.1% Valid Python 57.5% → 100% Peak VRAM 4.17 GB

The base model writes correct code — it just writes it in JavaScript about 40% of the time, wrapped in markdown fences. Pick a prompt:

Base LLaMA 3.2 3B 


    

    

      
Fine-tuned this adapter

      

    

  


  

    Outputs are pre-generated with greedy decoding, so they are deterministic and reproducible.
    Free HuggingFace Spaces have no GPU; running a 3B model live would take roughly 60 seconds
    per generation.


    Model and full evaluation:
    Raghul09/llama-code-gen-lora