felixmanojh/DJ-AI-Radio-LoRA

🎵 Tamil AI DJ Radio - LoRA Adapter

Fine-tuned Qwen 2.5-0.5B for generating energetic Tanglish (Tamil-English) radio DJ commentary

This is a LoRA adapter trained on 5,027 examples of Tanglish DJ commentary across 68 diverse themes. Perfect for researchers, fine-tuning enthusiasts, and Apple Silicon users with MLX.

---

🎯 Model Overview

• Base Model: Qwen/Qwen2.5-0.5B-Instruct (4-bit quantized)
• Adapter Type: LoRA (Low-Rank Adaptation)
• Training Data: 5,027 Tanglish DJ commentary examples
• Best Checkpoint: Iteration 2900 (validation loss: 1.856)
• Adapter Size: 17MB
• Framework: MLX (Apple Silicon optimized)

---

📊 Training Configuration

Parameter	Value
LoRA Rank	8
LoRA Alpha	16
Target Layers	16 attention layers
Dropout	0.05
Training Iterations	6,000 (best @ 2900)
Batch Size	4
Learning Rate	1e-4
Warmup Steps	100
Validation Loss	1.856
LoRA Parameters	4.4M (0.89% of base)

---

🚀 Quick Start

Installation

pip install mlx mlx-lm

Using with MLX

from mlx_lm import load, generate

# Load base model with LoRA adapter
model, tokenizer = load(
    "Qwen/Qwen2.5-0.5B-Instruct",
    adapter_path="felixmanojh/DJ-AI-Radio-LoRA"
)

# Create DJ commentary
messages = [
    {"role": "system", "content": "You are a Tamil AI radio DJ who speaks energetic Tanglish. Create engaging commentary."},
    {"role": "user", "content": "Hype up a high-energy dance track for weekend party"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=prompt, max_tokens=150, verbose=False)
print(response)

Example Output:

Party mode activate! Friday night ah Saturday night mode activate!
Club vibes high-energy vibes! Dance floor crowded! Everyone jumping!
Party starter! Energy maximum! Music energizing! Dancefloor energy!
Party anthem! Ready for any day!

Using with Transformers + PEFT

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-0.5B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(
    base_model,
    "felixmanojh/DJ-AI-Radio-LoRA"
)

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")

# Generate
messages = [
    {"role": "system", "content": "You are a Tamil AI radio DJ who speaks energetic Tanglish."},
    {"role": "user", "content": "Introduce a chill-out song"}
]

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=150)
print(tokenizer.decode(outputs[0]))

---

💡 Why Use the Adapter?

• 🪶 Lightweight: Only 17MB vs 276MB for full model
• 🔬 Research-Friendly: Easy to experiment with different base models
• 🚀 MLX Optimized: Blazing fast on Apple Silicon (M1/M2/M3/M4)
• 💰 Storage Efficient: Share multiple variants without duplicating base model
• 🔧 Customizable: Fine-tune further or merge with other adapters

---

📚 Training Data

Data Source

• Generation: Claude API (Anthropic)
• Batches: 68 diverse theme batches
• Total Examples: 5,027 training + 503 validation
• Format: Chat template (system + user prompt → DJ commentary)

Data Coverage

Category	Themes
Vibes	Party, Chill, Workout, Study, Morning, Night
Moods	Romantic, Nostalgic, Trending, Festival, Cultural Fusion
Activities	Road Trip, Beach, Rain, Commute, Gaming, Street Food
Genres	Dance, EDM, Indie, Retro, Remix, Acoustic
Audience	College, Family, Couples, Weekend, Late Night

Data Quality

• Diversity: 68 unique thematic batches
• Code-Mixing: Natural Tamil-English switching patterns
• Energy Levels: High-energy party to calm study vibes
• Cultural Context: Tamil cinema, festivals, street culture references

---

📈 Performance Metrics

Metric	Value
Validation Loss	1.856 (checkpoint 2900)
Average Output Length	~50-100 tokens
Coherence	High (selected for best quality vs checkpoint 1700)
Repetition	Low (avoided by checkpoint selection)
Energy Match	95%+ match to requested vibe

---

🎓 Intended Use

✅ Recommended

• Radio/podcast DJ commentary generation
• Content creation for Tamil music channels
• AI voice assistant for music apps
• Educational demonstrations of code-mixing
• Fine-tuning research on multilingual models

❌ Not Recommended

• Real-time conversation (use chat models instead)
• Factual information retrieval
• Long-form content generation
• Non-Tamil/English language pairs

---

🔄 Model Variants

Model	Format	Size	Use Case
DJ-AI-Radio-LoRA (this)	LoRA adapter	17MB	Research, MLX, training
DJ-AI-Radio	Merged (HF)	276MB	Deployment, Spaces, GPU
DJ-AI-Radio-MLX	Fused (MLX)	276MB	Apple Silicon inference

---

🌐 Live Demo

Try it live with voice cloning and AI music: Tamil AI DJ Radio Space

---

📝 Citation

@software{tamil_ai_dj_radio_2025,
  author = {Felix Manojh},
  title = {Tamil AI DJ Radio - Tanglish Commentary Generator},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/felixmanojh/DJ-AI-Radio-LoRA},
  note = {LoRA adapter for Qwen 2.5-0.5B with Claude-generated training data}
}

---

📄 License

Apache 2.0 (inherits from Qwen 2.5)

Free for commercial and non-commercial use.

---

🙏 Acknowledgments

• Base Model: Qwen Team (Alibaba Cloud)
• Training Framework: MLX (Apple)
• Training Data: Generated with Claude API (Anthropic)
• Inspiration: Tamil music culture and radio DJ traditions

---

Built with ❤️ for the Tamil-speaking community