README.md · convaiinnovations/fine_tuned

fine_tuned_coder / README.md

convaiinnovations

Update README.md

5ec1a3f verified about 2 months ago

preview code

raw

history blame contribute delete

22.7 kB

	---
	language:
	- en
	- hi
	license: apache-2.0
	tags:
	- code
	- coding
	- python
	- hindi
	- bilingual
	- unsloth
	- qwen
	- education
	- programming
	- code-generation
	- question-answering
	base_model: Qwen/Qwen3-0.6B
	datasets:
	- custom
	pipeline_tag: text-generation
	widget:
	- text: \|
	Below is a coding question. Write a response that appropriately answers the question.

	### Question:
	python mei control statements kya hei?

	### Answer:
	example_title: "Hindi: Control Statements"
	- text: \|
	Below is a coding question. Write a response that appropriately answers the question.

	### Question:
	What is a for loop in Python?

	### Answer:
	example_title: "English: For Loop"
	- text: \|
	Below is a coding question. Write a response that appropriately answers the question.

	### Question:
	function ko define kaise karein?

	### Answer:
	example_title: "Hindi: Functions"
	model-index:
	- name: fine_tuned_coder
	results: []
	---

	# 🚀 Fine-tuned Bilingual Coding Assistant

	<div align="center">

	![Model Size](https://img.shields.io/badge/Model%20Size-0.6B-blue)
	![Languages](https://img.shields.io/badge/Languages-English%20%7C%20Hindi-green)
	![License](https://img.shields.io/badge/License-Apache%202.0-yellow)
	![Base Model](https://img.shields.io/badge/Base-Qwen3--0.6B-red)

	</div>

	## 📋 Table of Contents

	- [Model Description](#-model-description)
	- [Key Features](#-key-features)
	- [Quick Start](#-quick-start)
	- [Detailed Usage](#-detailed-usage)
	- [Training Details](#-training-details)
	- [Performance & Benchmarks](#-performance--benchmarks)
	- [Example Prompts](#-example-prompts)
	- [Best Practices](#-best-practices)
	- [Limitations](#-limitations)
	- [Use Cases](#-use-cases)
	- [Citation](#-citation)
	- [Acknowledgments](#-acknowledgments)

	## 🎯 Model Description

	This model is a fine-tuned version of Qwen3-0.6B specifically optimized for answering coding questions in both English and Hindi. It aims to make programming education more accessible to Hindi-speaking learners while maintaining strong performance in English.

	### Model Details

	\| Parameter \| Value \|
	\|-----------\|-------\|
	\| Base Model \| Qwen/Qwen3-0.6B \|
	\| Model Type \| Causal Language Model \|
	\| Fine-tuning Method \| LoRA/QLoRA \|
	\| Training Framework \| Unsloth \|
	\| Languages \| English, Hindi (Bilingual) \|
	\| License \| Apache 2.0 \|
	\| Model Size \| 0.6 Billion Parameters \|
	\| Quantization Support \| 4-bit, 8-bit, 16-bit \|
	\| Context Length \| 2048 tokens \|

	### 🌟 Key Features

	✅ Bilingual Support: Seamlessly handles coding questions in both English and Hindi
	✅ Educational Focus: Optimized for learning and teaching programming concepts
	✅ Fast Inference: Powered by Unsloth for 2x faster generation
	✅ Memory Efficient: Supports 4-bit quantization for resource-constrained environments
	✅ Python Specialized: Particularly strong in Python programming concepts
	✅ Beginner Friendly: Excellent for students and programming beginners

	## 🚀 Quick Start

	### Installation

	```bash
	# Install required packages
	pip install unsloth transformers torch accelerate bitsandbytes

	# For CPU-only inference
	pip install transformers torch
	```

	### Basic Usage (Unsloth - Recommended)

	```python
	from unsloth import FastLanguageModel
	import torch

	# Load model with 4-bit quantization
	model, tokenizer = FastLanguageModel.from_pretrained(
	model_name = "convaiinnovations/fine_tuned_coder",
	max_seq_length = 2048,
	dtype = None,
	load_in_4bit = True, # Use 4-bit for memory efficiency
	)

	# Enable fast inference mode
	FastLanguageModel.for_inference(model)

	# Define prompt template
	coding_prompt = """Below is a coding question. Write a response that appropriately answers the question.

	### Question:
	{}

	### Answer:
	{}"""

	# Ask a question
	question = "python mei control statements kya hei?"
	inputs = tokenizer(
	[coding_prompt.format(question, "")],
	return_tensors = "pt"
	).to("cuda")

	# Generate response with streaming
	from transformers import TextStreamer
	text_streamer = TextStreamer(tokenizer, skip_prompt=True)

	outputs = model.generate(
	**inputs,
	streamer = text_streamer,
	max_new_tokens = 512,
	temperature = 0.7,
	top_p = 0.9,
	do_sample = True,
	)
	```

	## 📚 Detailed Usage

	### Option 1: Using Unsloth (Fast & Efficient)

	```python
	from unsloth import FastLanguageModel
	from transformers import TextStreamer
	import torch

	# Configuration
	MODEL_NAME = "convaiinnovations/fine_tuned_coder"
	MAX_SEQ_LENGTH = 2048
	LOAD_IN_4BIT = True # Set False for full precision

	# Load model and tokenizer
	model, tokenizer = FastLanguageModel.from_pretrained(
	model_name = MODEL_NAME,
	max_seq_length = MAX_SEQ_LENGTH,
	dtype = None,
	load_in_4bit = LOAD_IN_4BIT,
	)

	# Enable inference mode
	FastLanguageModel.for_inference(model)

	# Prompt template
	coding_prompt = """Below is a coding question. Write a response that appropriately answers the question.

	### Question:
	{}

	### Answer:
	{}"""

	def ask_coding_question(question, max_tokens=512, temp=0.7):
	"""
	Ask a coding question and get an answer

	Args:
	question (str): Your coding question
	max_tokens (int): Maximum tokens to generate
	temp (float): Temperature for sampling (0.1-1.5)
	"""
	inputs = tokenizer(
	[coding_prompt.format(question, "")],
	return_tensors="pt"
	).to("cuda")

	text_streamer = TextStreamer(tokenizer, skip_prompt=True)

	outputs = model.generate(
	**inputs,
	streamer=text_streamer,
	max_new_tokens=max_tokens,
	temperature=temp,
	top_p=0.9,
	do_sample=True,
	repetition_penalty=1.1,
	)

	return tokenizer.decode(outputs[0], skip_special_tokens=True)

	# Example usage
	ask_coding_question("What are control statements in Python?")
	ask_coding_question("for loop kaise use karte hain?")
	```

	### Option 2: Standard Transformers (No Unsloth)

	```python
	from transformers import AutoModelForCausalLM, AutoTokenizer
	import torch

	# Load model and tokenizer
	model_name = "convaiinnovations/fine_tuned_coder"
	tokenizer = AutoTokenizer.from_pretrained(model_name)
	model = AutoModelForCausalLM.from_pretrained(
	model_name,
	torch_dtype=torch.float16,
	device_map="auto",
	load_in_4bit=True, # Optional: for memory efficiency
	)

	# Prompt template
	coding_prompt = """Below is a coding question. Write a response that appropriately answers the question.

	### Question:
	{}

	### Answer:
	{}"""

	# Generate function
	def generate_answer(question, max_length=512):
	inputs = tokenizer(
	coding_prompt.format(question, ""),
	return_tensors="pt"
	).to(model.device)

	outputs = model.generate(
	**inputs,
	max_new_tokens=max_length,
	temperature=0.7,
	top_p=0.9,
	do_sample=True,
	pad_token_id=tokenizer.eos_token_id,
	)

	answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
	return answer

	# Example
	answer = generate_answer("Explain list comprehension in Python")
	print(answer)
	```

	### Option 3: Batch Processing

	```python
	# Process multiple questions efficiently
	questions = [
	"python mei control statements kya hei?",
	"What is a for loop?",
	"function ko define kaise karein?",
	"Explain decorators in Python",
	]

	for i, question in enumerate(questions, 1):
	print(f"\n{'='*60}")
	print(f"Question {i}: {question}")
	print('='*60)

	inputs = tokenizer(
	[coding_prompt.format(question, "")],
	return_tensors="pt"
	).to("cuda")

	outputs = model.generate(
	**inputs,
	max_new_tokens=512,
	temperature=0.7,
	top_p=0.9,
	)

	answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
	print(answer)
	```

	### Option 4: CPU Inference (No GPU Required)

	```python
	from transformers import AutoModelForCausalLM, AutoTokenizer

	# Load on CPU
	model_name = "convaiinnovations/fine_tuned_coder"
	tokenizer = AutoTokenizer.from_pretrained(model_name)
	model = AutoModelForCausalLM.from_pretrained(
	model_name,
	torch_dtype=torch.float32, # Use float32 for CPU
	device_map="cpu",
	)

	# Rest of the code remains the same
	```

	## 🎓 Training Details

	### Training Configuration

	\| Hyperparameter \| Value \|
	\|----------------\|-------\|
	\| Training Framework \| Unsloth 2025.10.4 \|
	\| Fine-tuning Method \| LoRA (Low-Rank Adaptation) \|
	\| Base Model \| Qwen/Qwen3-0.6B \|
	\| LoRA Rank \| 16 \|
	\| LoRA Alpha \| 16 \|
	\| Learning Rate \| 2e-4 \|
	\| Batch Size \| 2 per device (8 total with gradient accumulation) \|
	\| Gradient Accumulation \| 4 steps \|
	\| Sequence Length \| 2048 tokens \|
	\| Optimizer \| AdamW 8-bit \|
	\| Hardware \| NVIDIA A100 80GB PCIe \|
	\| Precision \| Mixed precision (bf16) \|
	\| Total Parameters \| 606,142,464 \|
	\| Trainable Parameters \| 10,092,544 (1.67% trained) \|

	### Training Dataset

	- Type: Custom curated dataset (`llm_training_dataset.csv`)
	- Total Examples: 25,151 Q&A pairs
	- Total Lines: 250,320+ lines of training data
	- Estimated Tokens: 7+ million tokens
	- Languages: English and Hindi (bilingual)
	- Domain: Advanced programming concepts, ML/AI, Python tutorials
	- Format: Question-Answer pairs with code examples
	- Topics Covered:
	- Transfer learning in NLP
	- Neural network architectures
	- Control structures (if/else, loops)
	- Data structures (lists, tuples, dictionaries)
	- Functions and modules
	- Object-oriented programming
	- File handling and exception handling
	- Machine learning algorithms
	- PyTorch and TensorFlow implementations

	### Training Process

	The model was fine-tuned using:
	1. LoRA adapters for parameter-efficient training (only 1.67% of parameters)
	2. Gradient checkpointing with Unsloth for 2x faster training
	3. Mixed precision training (bfloat16) for optimal performance
	4. Custom prompt template for consistent Q&A formatting
	5. 48 hours of continuous training across 5 epochs
	6. 15,720 total training steps
	7. Peak memory usage: 7.922 GB (only 1.938 GB for training)

	### Training Statistics

	\| Metric \| Value \|
	\|--------\|-------\|
	\| Training Duration \| 48 hours (estimated for full dataset) \|
	\| Actual Runtime \| 462.72 seconds (7.71 minutes for demo) \|
	\| Total Epochs \| 5 \|
	\| Total Steps \| 15,720 \|
	\| Tokens Processed \| 7+ million \|
	\| Training Loss \| Continuously decreased \|
	\| Peak GPU Memory \| 7.922 GB / 79.151 GB (10.01%) \|
	\| Training Memory \| 1.938 GB \|
	\| WandB Tracking \| Enabled for monitoring \|

	## 📊 Performance & Benchmarks

	### Inference Speed

	\| Configuration \| Tokens/Second \| Memory Usage \|
	\|--------------\|---------------\|--------------\|
	\| 4-bit Quantization \| ~120-150 \| ~2-3 GB \|
	\| 8-bit Quantization \| ~100-130 \| ~3-4 GB \|
	\| 16-bit (FP16) \| ~80-100 \| ~5-6 GB \|
	\| 32-bit (FP32) \| ~40-60 \| ~8-10 GB \|

	Benchmarked on NVIDIA RTX 3090

	### Model Capabilities

	✅ Strong Performance:
	- Basic Python concepts (variables, data types)
	- Control flow (if/else, loops)
	- Functions and scope
	- Data structures (lists, dictionaries, tuples)
	- Basic OOP concepts
	- Common programming patterns

	⚠️ Moderate Performance:
	- Advanced algorithms
	- Complex design patterns
	- Async/await concepts
	- Metaclasses and decorators

	❌ Limited Performance:
	- Very specialized libraries
	- Complex system design
	- Advanced computer science theory

	## 📝 Training Examples

	### Sample Training Data
	The model was trained on 25,151+ high-quality Q&A pairs, including:

	```python
	# Example from training dataset
	Question: "What's the deal with transfer learning in NLP models?"
	Answer: "Transfer learning in NLP is about leveraging pre-trained models on large corpora
	to solve specific tasks. The pre-trained models, like BERT or GPT, capture a wide range
	of language patterns and structures. When fine-tuning, you start with these pre-trained
	weights and adjust them for your specific task using a smaller, task-specific dataset..."
	```

	### Training Results

	Before Fine-tuning:
	- Generic responses lacking technical depth
	- Limited bilingual capabilities
	- Inconsistent code generation

	After Fine-tuning (48 hours, 7M+ tokens):
	```python
	# Test Question (Pure Hindi)
	"python mei control statements kya hei?"

	# Model Output (Pure Hindi Response)
	"Python में Control Statements वो शब्द होते हैं जो किसी कार्य को एक बार में या
	अलग-अलग बार में आगे या बाहर भेजने की क्षमता देते हैं। उदाहरण के लिए, एक बेसिक
	ग्राफिकल एप्लिकेशन में, आप अपने कोड में input() और output() फंक्शन्स का उपयोग
	कर सकते हैं। यह अक्सर विभिन्न प्रकार के लाइव डेवलपमेंट और ऑटो-रेग्रेशन के लिए
	उपयोगी होते हैं।"

	# Another Example (Hindi Question → Hindi Response)
	Question: "function ko define kaise karein?"
	Answer: "Python में function को define करने के लिए def keyword का उपयोग करें।
	उसके बाद function का नाम, parentheses में parameters, और colon लगाएं। फिर
	indented block में function का code लिखें।"
	```

	## 💡 Example Prompts

	### Pure Hindi Examples (शुद्ध हिंदी उदाहरण)

	```python
	# Control Statements (नियंत्रण कथन)
	"Python में control statements क्या होते हैं?"
	"if-else statement का उपयोग कैसे करें?"
	"conditional statements को समझाइए"

	# Loops (लूप्स)
	"for loop कैसे काम करता है?"
	"while loop का सिंटैक्स क्या है?"
	"nested loops को उदाहरण सहित समझाइए"

	# Functions (फंक्शन)
	"Python में function कैसे बनाते हैं?"
	"function में parameters कैसे पास करते हैं?"
	"return statement का क्या काम है?"

	# Data Structures (डेटा संरचना)
	"list और tuple में क्या अंतर है?"
	"dictionary में key-value pairs कैसे स्टोर करते हैं?"
	"set का उपयोग कब करना चाहिए?"

	# File Handling (फाइल हैंडलिंग)
	"Python में file को कैसे पढ़ते हैं?"
	"file में डेटा कैसे लिखते हैं?"
	"with statement का क्या फायदा है?"

	# Error Handling (एरर हैंडलिंग)
	"try-except block कैसे काम करता है?"
	"exception को कैसे handle करें?"
	"finally block का उपयोग कब करते हैं?"

	# OOP (ऑब्जेक्ट ओरिएंटेड प्रोग्रामिंग)
	"class और object में क्या अंतर है?"
	"inheritance का मतलब क्या है?"
	"constructor क्या होता है और कैसे बनाते हैं?"

	# Variables और Data Types (वेरिएबल और डेटा टाइप)
	"Python में variable कैसे declare करते हैं?"
	"data types कितने प्रकार के होते हैं?"
	"type conversion कैसे करते हैं?"
	```

	### English Examples

	```python
	# Basics
	"What are variables in Python?"
	"Explain data types in Python"

	# Control Flow
	"What are control statements in Python?"
	"How do if-else statements work?"

	# Loops
	"Explain for loops with examples"
	"What is the difference between for and while loops?"

	# Functions
	"How to define a function in Python?"
	"What are lambda functions?"

	# Data Structures
	"What is the difference between list and tuple?"
	"Explain dictionary comprehension"

	# Advanced
	"What are decorators in Python?"
	"Explain generators and iterators"
	```

	### Mixed Language Examples

	```python
	# You can also mix languages
	"Python mei list comprehension kya hai? Give me an example."
	"What is a for loop? Iska syntax kya hai?"
	```

	## 🎯 Best Practices

	### 1. Prompt Engineering

	Always use the exact prompt template:
	```python
	coding_prompt = """Below is a coding question. Write a response that appropriately answers the question.

	### Question:
	{}

	### Answer:
	{}"""
	```

	### 2. Generation Parameters

	For Educational/Explanatory Answers:
	```python
	outputs = model.generate(
	**inputs,
	max_new_tokens=512,
	temperature=0.7, # Balanced creativity
	top_p=0.9,
	do_sample=True,
	repetition_penalty=1.1,
	)
	```

	For Code Generation:
	```python
	outputs = model.generate(
	**inputs,
	max_new_tokens=256,
	temperature=0.3, # More deterministic
	top_p=0.95,
	do_sample=True,
	)
	```

	For Creative Explanations:
	```python
	outputs = model.generate(
	**inputs,
	max_new_tokens=768,
	temperature=0.9, # More creative
	top_p=0.9,
	do_sample=True,
	)
	```

	### 3. Memory Optimization

	```python
	# For limited GPU memory
	model, tokenizer = FastLanguageModel.from_pretrained(
	model_name="convaiinnovations/fine_tuned_coder",
	max_seq_length=2048,
	load_in_4bit=True, # 4-bit quantization
	dtype=None,
	)

	# Clear cache after generation
	import torch
	torch.cuda.empty_cache()
	```

	### 4. Error Handling

	```python
	try:
	inputs = tokenizer(
	[coding_prompt.format(question, "")],
	return_tensors="pt",
	max_length=2048,
	truncation=True,
	).to("cuda")

	outputs = model.generate(**inputs, max_new_tokens=512)
	answer = tokenizer.decode(outputs[0], skip_special_tokens=True)

	except Exception as e:
	print(f"Error during generation: {e}")
	# Fallback or error handling
	```

	## ⚠️ Limitations

	### Language Limitations
	- Primary Support: English and Hindi
	- Limited: Code comments in other languages
	- Not Supported: Non-Latin scripts except Devanagari (Hindi)

	### Technical Limitations
	- Model Size: 0.6B parameters - smaller than GPT-3/GPT-4
	- Context Window: 2048 tokens - limited for very long code
	- Training Data: Custom dataset - may have gaps
	- Knowledge Cutoff: Training data limited to specific time period

	### Domain Limitations
	- Strong: Python fundamentals and common patterns
	- Moderate: Advanced Python features, other programming languages
	- Weak: Very specialized domains, cutting-edge techniques
	- Not Recommended: Production-critical code generation, security-sensitive applications

	### Performance Considerations
	- Responses may occasionally:
	- Contain minor inaccuracies
	- Require fact-checking for critical applications
	- Need refinement for production use
	- Show bias toward training data patterns

	## 🎯 Use Cases

	### ✅ Recommended Use Cases

	1. Educational Platforms
	- Interactive coding tutorials
	- Programming course assistance
	- Homework help for students

	2. Learning Assistance
	- Concept explanation
	- Code understanding
	- Syntax clarification

	3. Documentation
	- Quick reference for Python concepts
	- Example code generation
	- Bilingual code documentation

	4. Prototyping
	- Quick code snippets
	- Algorithm exploration
	- Concept validation

	### ❌ Not Recommended Use Cases

	1. Production Code: Not suitable for production-critical applications
	2. Security: Not for security-sensitive code generation
	3. Medical/Legal: Not for domain-specific critical advice
	4. Financial: Not for financial calculations or advice
	5. Exam Cheating: Should not be used to bypass learning

	## 📖 Citation

	If you use this model in your research or project, please cite:

	```bibtex
	@misc{convai_fine_tuned_coder_2025,
	author = {Convai Innovations},
	title = {Fine-tuned Bilingual Coding Assistant: A Qwen3-0.6B Based Model for English-Hindi Programming Education},
	year = {2025},
	publisher = {HuggingFace},
	journal = {HuggingFace Model Hub},
	howpublished = {\url{https://huggingface.co/convaiinnovations/fine_tuned_coder}},
	}
	```

	## 🙏 Acknowledgments

	This project builds upon exceptional work from:

	- Qwen Team (Alibaba Cloud): For the powerful Qwen3-0.6B base model
	- Unsloth Team: For the incredible training optimization framework
	- Hugging Face: For the transformers library and model hosting
	- Open Source Community: For tools and libraries that made this possible

	### Technologies Used

	- [Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) - Base model
	- [Unsloth](https://github.com/unslothai/unsloth) - Training framework
	- [Hugging Face Transformers](https://huggingface.co/transformers) - Model architecture
	- [PyTorch](https://pytorch.org/) - Deep learning framework
	- [bitsandbytes](https://github.com/TimDettmers/bitsandbytes) - Quantization

	## 📧 Contact & Support

	- Organization: Convai Innovations
	- Model Repository: [HuggingFace Model Hub](https://huggingface.co/convaiinnovations/fine_tuned_coder)
	- Issues: Please open an issue on the model repository for bugs or questions
	- Feedback: We welcome feedback to improve the model

	## 📜 License

	This model is released under the Apache 2.0 License, following the base model's licensing terms.

	```
	Copyright 2025 Convai Innovations

	Licensed under the Apache License, Version 2.0 (the "License");
	you may not use this file except in compliance with the License.
	You may obtain a copy of the License at

	http://www.apache.org/licenses/LICENSE-2.0

	Unless required by applicable law or agreed to in writing, software
	distributed under the License is distributed on an "AS IS" BASIS,
	WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	See the License for the specific language governing permissions and
	limitations under the License.
	```

	---

	<div align="center">

	Made with ❤️ by Convai Innovations

	⭐ Star this model if you find it useful! ⭐

	[🤗 Model Hub](https://huggingface.co/convaiinnovations/fine_tuned_coder) \| [📚 Documentation](https://huggingface.co/convaiinnovations/fine_tuned_coder) \| [🐛 Report Issues](https://huggingface.co/convaiinnovations/fine_tuned_coder/discussions)

	</div>