metadata
license: other
license_name: codet5-salesforce
license_link: LICENSE
language:
- en
base_model:
- Salesforce/codet5-small
pipeline_tag: text-generation
tags:
- code
- commit
- gitdiff
datasets:
- seniruk/git-diff_to_commit_msg_large
Fintuned Salesforce/codet5-small base model using 1000000 rows of data with git commits of different types of random languages for 5 epochs
Took a total of 10 hours in a gpu of RTX 4060TI 16GB VRAM.
Use the below instructions for inference
Modules required- transformers,pytorch,CUDA
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
Load the correct CodeT5 tokenizer and model
model_name = "Salesforce/codet5-small"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
#Example Git diff input
git_diff = """
diff --git a/example.py b/example.py
index 3b18e12..b3f7e54 100644
--- a/example.py
+++ b/example.py
@@ -1,5 +1,6 @@
-def greet():
- print("Hello, world!")
+def greet_user(name):
+ print(f"Hello, {name}!")
-def farewell():
- print("Goodbye!")
+def farewell_user(name):
+ print(f"Goodbye, {name}!")
"""
#keep the instruction unchanged, becus the model was trained on this static instruction
instruction = "Generate a commit message based on the following Git diff:\n\n"
task_input = instruction + git_diff
# Tokenize the input
inputs = tokenizer(
task_input,
max_length=512, # Truncate if necessary
truncation=True,
padding="max_length",
return_tensors="pt"
)
# Generate commit message
outputs = model.generate(
inputs["input_ids"],
max_length=50,
num_beams=5, # Use beam search
temperature=0.9, # Adds controlled randomness
top_p=0.9, # Nucleus sampling
early_stopping=True
)
# Decode the generated commit message
commit_message = tokenizer.decode(outputs[0], skip_special_tokens=True)
# Print the result
print("Generated Commit Message:")
print(commit_message)