metadata
license: other
license_name: seniru-epasinghe
license_link: LICENSE
language:
- en
base_model:
- Qwen/Qwen2.5-Coder-0.5B
pipeline_tag: text-generation
datasets:
- seniruk/git-diff_to_commit_msg_large
Finetuned-qwen2.5-coder-0.5B model on 100000 rows of a cutom dataset containing. git-differences and respective commit messages
Each row of the dataset was formatted as below to suit finetuning requirement of Qwen2.5-coder model
'### Instruction:\nGenerate a concise and meaningful commit message based on the provided git diff.\n\n### Git Diff:\n{a given git-difference as in the dataset rows}\n\n### Commit Message:\nAdding the squeezing in the cost fuction<|im_end|>'
Code for inference of the gguf model is given below
from llama_cpp import Llama
# Configuration
gguf_model_path = "qwen0.5-finetuned.gguf" # Path to your GGUF file
# Define the commit message prompt (Minimal format, avoids assistant behavior)
commit_prompt = """Generate a meaningful commit message explaining all the changes in the provided Git diff.
### Git Diff:
{}
### Commit Message:""" # Removed {} after "Commit Message:" to prevent pre-filled text.
# Git diff example for commit message generation
git_diff_example = """
diff --git a/index.html b/index.html
index 89abcde..f123456 100644
--- a/index.html
+++ b/index.html
@@ -5,16 +5,6 @@ <body>
<h1>Welcome to My Page</h1>
- <table border="1">
- <tr>
- <th>Name</th>
- <th>Age</th>
- </tr>
- <tr>
- <td>John Doe</td>
- <td>30</td>
- </tr>
- </table>
+ <p>This is a newly added paragraph replacing the table.</p>
</body>
</html>
"""
# Load the GGUF model with increased context size (32768)
modelGGUF = Llama(
model_path=gguf_model_path,
rope_scaling={"type": "linear", "factor": 2.0},
chat_format=None, # Disables any chat formatting
n_ctx=32768, # Set the context size explicitly
)
# Prepare the raw input prompt
input_prompt = commit_prompt.format(git_diff_example)
# Generate commit message
output = modelGGUF(
input_prompt,
max_tokens=64,
temperature=0.6, # Balanced randomness
top_p=0.8, # Controls nucleus sampling
top_k=50, # Limits vocabulary selection
)
# Decode and print the output
commit_message = output["choices"][0]["text"].strip()
print("\nGenerated Commit Message:\n{}".format(commit_message))