Qwen2.5-Coder-3B-Instruct – Fine-tuned for COBOL Code Documentation

This model is a fine-tuned version of Qwen/Qwen2.5-Coder-3B-Instruct, optimized for generating natural language documentation from COBOL source code. The fine-tuning was done using freeze fine-tuning on the last transformer layer only, preserving the rest of the model's pretrained weights.

🔧 Model Description

Architecture: Qwen2.5-Coder-3B (decoder-only transformer)
Base Model: Qwen/Qwen2.5-Coder-3B-Instruct
Fine-tuning Method: Freeze fine-tuning (only last transformer block's parameters were updated)
Training Objective: Instruction-following text generation for COBOL code documentation

🧠 Use Cases

This model is specialized in generating descriptive documentation for legacy COBOL code, especially useful for:

Legacy system maintenance
Automated codebase documentation
Migration planning
COBOL code understanding and onboarding

✍️ Example Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

model_name = "V7W3D/qwen-code-doc-ft"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

doc_gen = pipeline("text-generation", model=model, tokenizer=tokenizer)

prompt = "### Document this COBOL code:\n\n       IDENTIFICATION DIVISION.\n       PROGRAM-ID. HELLO-WORLD.\n       PROCEDURE DIVISION.\n           DISPLAY 'HELLO, WORLD!'\n           STOP RUN.\n\n### Documentation:"
response = doc_gen(prompt, max_new_tokens=200, do_sample=False)

print(response[0]["generated_text"])