Saikrishna2511/qwen-multitask

Saikrishna2511/qwen-multitask

Multi-task fine-tuned Qwen2.5-Coder-0.5B-Instruct checkpoint for code generation and documentation.

Demo

Try the model in the browser: https://huggingface.co/spaces/Saikrishna2511/qwen-multitask-demo

Tasks

This single checkpoint handles three tasks via different prompt prefixes:

NL → Python (nl2py)

### Instruction: Write Python for: {natural language description}
### Response:

Java → Python (java2py)

### Translate Java to Python:

{java code}

### Python:

### Code → Documentation (`code2doc`)

Generate documentation for this Python code:

{python code}

Documentation:

## Training

- **Base model:** [Qwen/Qwen2.5-Coder-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-0.5B-Instruct)
- **Stage 1:** Java→Python LoRA fine-tune on AVATAR-TC
- **Stage 2:** Multi-task LoRA on NL2Py, Code2Doc, code comments, and Java2Py replay
- **Method:** LoRA (r=16, alpha=32), merged weights for inference

## Usage

from transformers import AutoModelForCausalLM, AutoTokenizer import torch

modelid = "Saikrishna2511/qwen-multitask" tokenizer = AutoTokenizer.frompretrained(modelid, trustremotecode=True) model = AutoModelForCausalLM.frompretrained( modelid, trustremotecode=True, torchdtype=torch.float16, device_map="auto", )

prompt = "### Instruction: Write Python for: return the factorial of n\n### Response:\n" inputs = tokenizer(prompt, returntensors="pt").to(model.device) outputs = model.generate(**inputs, maxnewtokens=512, temperature=0.2, topp=0.95) print(tokenizer.decode(outputs[0], skipspecialtokens=True))

For post-processing and all three task templates, see the [project repo](https://github.com) or the linked Gradio Space.

## Limitations

- Small 0.5B model; quality varies by task and input complexity
- Trained primarily on Python; Java translation quality depends on training coverage
- Not intended for production use without further evaluation