---

license: apache-2.0

datasets:

• sahil2801/CodeAlpaca-20k

language:

• en

base_model:

• google/gemma-3-270m-it

new_version: Rev2auth/gemma3-270m-codealpaca-lora

---

Gemma 3 270M CodeAlpaca LoRA

A LoRA fine-tuned version of Google Gemma 3 270M Instruct for Python programming and code generation tasks.

Model Information

• Base Model: google/gemma-3-270m-it
• Model Type: LoRA Adapter
• Framework: PEFT (LoRA)
• Language: English
• Primary Task: Code Generation
• Training Dataset: CodeAlpaca

Overview

This project fine-tunes Gemma 3 270M on the CodeAlpaca dataset to improve its ability to generate, complete, and explain Python code.

The model is designed for:

• Python code generation
• Function completion
• Code explanation
• Basic debugging assistance
• Programming question answering

Training

Dataset

• CodeAlpaca

Fine-Tuning Method

• LoRA (Low-Rank Adaptation)

Base Model

• google/gemma-3-270m-it

Estimated Performance

> Note: The following values are subjective estimates based on manual testing and are not official benchmark scores.

| Metric | Estimated Score |

|---------|----------------:|

| Python Syntax | 90% |

| Function Structure | 85% |

| Instruction Following | 70% |

| Code Completion | 65% |

| Algorithm Correctness | 40% |

| Overall Coding Ability | 70% |

Estimated benchmark equivalents:

• HumanEval pass@1: 8–12%
• MBPP pass@1: 18–25%

Example

Prompt

Write a Python function to reverse a linked list.

Response

def reverse_linked_list(head):
    prev = None
    current = head

    while current:
        nxt = current.next
        current.next = prev
        prev = current
        current = nxt

    return prev

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "google/gemma-3-270m-it"
)

model = PeftModel.from_pretrained(
    base_model,
    "Rev2auth/gemma3-270m-codealpaca-lora"
)

tokenizer = AutoTokenizer.from_pretrained(
    "google/gemma-3-270m-it"
)

Limitations

• Small 270M parameter model
• Limited algorithmic reasoning
• May generate logically incorrect solutions
• Best suited for beginner and intermediate coding tasks

Future Work

• Improve benchmark performance
• Train on higher-quality coding datasets
• Add multi-language programming support
• Increase algorithmic reasoning ability
• Release improved versions

Repository

LoRA Adapter:

https://huggingface.co/Rev2auth/gemma3-270m-codealpaca-lora

GGUF Model:

https://huggingface.co/Rev2auth/gemma3-270m-codealpaca-gguf

License

Please follow the license of the original Gemma 3 base model.

Acknowledgements

• Google DeepMind for Gemma 3
• Hugging Face
• PEFT
• TRL
• CodeAlpaca Dataset

---

Author: Revanth Annapureddy

Version: v1.0

Status: First public fine-tuned release