metadata
library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen2.5-1.5B-Instruct
datasets:
- iamtarun/python_code_instructions_18k_alpaca
language:
- ar
- en
pipeline_tag: text-generation
tags:
- llama-factory
- lora
- qwen2
- python
- arabic
- code
- instruction-tuning
- fine-tuned
🐍 Python Assistant (Arabic)
A fine-tuned version of Qwen2.5-1.5B-Instruct that answers Python programming questions in Arabic, with structured JSON output. Fine-tuned using LoRA via LLaMA-Factory.
Model Details
- Developed by: jana-ashraf-ai
- Base Model: Qwen/Qwen2.5-1.5B-Instruct
- Model type: Causal Language Model (text-generation)
- Language(s): Arabic (answers) + English (questions)
- License: Apache 2.0
- Fine-tuning method: QLoRA (LoRA rank=32) via LLaMA-Factory
What does this model do?
Given a Python programming question in English, the model returns a structured JSON answer in Arabic, explaining the solution step by step.
How to Use
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_id = "jana-ashraf-ai/python-assistant"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.float16,
device_map="auto"
)
system_prompt = """You are a Python expert assistant.
Answer the user's Python question in Arabic following the Output Schema.
Do not add any introduction or conclusion."""
question = "How do I reverse a list in Python?"
messages = [
{"role": "system", "content": system_prompt},
{"role": "user", "content": question}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Training Details
| Parameter | Value |
|---|---|
| Base model | Qwen2.5-1.5B-Instruct |
| Fine-tuning method | LoRA (QLoRA) |
| LoRA rank | 32 |
| LoRA target | all |
| Training samples | 1,000 |
| Epochs | 3 |
| Learning rate | 1e-4 |
| LR scheduler | cosine |
| Warmup ratio | 0.1 |
| Batch size | 1 (grad accum = 8) |
| Precision | fp16 |
| Quantization | 4-bit (nf4) |
| Framework | LLaMA-Factory |
| Hardware | Google Colab T4 GPU |
Training Data
Fine-tuned on a curated subset (1,000 samples) from iamtarun/python_code_instructions_18k_alpaca.
The answers were annotated and structured using GPT to produce Arabic explanations in a JSON schema format.
Train / Val split: 90% / 10%
Limitations
- The model is optimized for Python questions only.
- Answers are in Arabic — not suitable for English-only use cases.
- Small model size (1.5B) may struggle with very complex programming problems.
- Output quality depends on the question being clear and specific.