QLoRA-Finetuning

Llama-3.2-3B · CodeAlpaca LoRA Adapter

A LoRA adapter fine-tuned on CodeAlpaca-20k for instruction-following code generation tasks. Built on top of meta-llama/Llama-3.2-3B with 4-bit NF4 quantization via bitsandbytes. Only ~1% of parameters are trainable — the rest of the base model is frozen.

---

Model Details

---

Training Configuration

---

Dataset

• Name: CodeAlpaca-20k
• Size: ~20,000 code instruction samples
• Split: 90/10 train/test (~18,000 train, ~2,000 test)
• Columns: instruction, input, output
• Prompt format: Instruction: {instruction}

Input: {input}

Response: {output}

text

---

Evaluation Results

Evaluated on 200 held-out test samples from CodeAlpaca-20k using 4-bit quantized inference. Metrics computed with evaluate (ROUGE-L) and bert_score (BERTScore-F1).

ROUGE-L of 0.5458 is at the top of the competitive range for fine-tuned code generation models (0.43–0.55), confirming that LoRA fine-tuning successfully taught the model consistent instruction-following and code formatting behavior.

---

How to Use

Load the base model with 4-bit quantization, then apply this adapter using PEFT's PeftModel.from_pretrained().

Prompt format: Instruction: Write a Python function that reverses a string.

Input: Response: text

Inference parameters used during evaluation:

• max_new_tokens: 200
• do_sample: False
• repetition_penalty: 1.1
• pad_token_id: tokenizer.eos_token_id

---

Limitations

• Trained for only 1–3 epochs on 18k samples — may struggle with highly complex or multi-file code tasks.
• Optimized for single-instruction, single-response code generation; not designed for multi-turn conversation.
• Performance is measured on CodeAlpaca-style prompts; may degrade on very different prompt formats.
• Base model is 3B parameters — larger models (7B+) would likely achieve higher absolute scores.

---

Project

This adapter was built as part of a 7-day end-to-end LLM fine-tuning project covering LoRA/QLoRA concepts, dataset preparation, training, evaluation, deployment, and CI/CD. Full project repository: github.com/your-username/llm-lora-finetuning