Learning Rate: 2e-4
Batch Size: 4 (effective: 16 with gradient accumulation)
Gradient Accumulation Steps: 4
LoRA Rank (r): 16
LoRA Alpha: 32
LoRA Dropout: 0.05
Optimizer: paged_adamw_32bit
Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Trainable Parameters: 41,943,040 (0.52% of total)

🚀 Quick Start

python

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

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3-8B-Instruct",
    device_map="auto",
    torch_dtype=torch.float16
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "owenergy/llama3-sharegpt-10k-voice-ai")

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct")

# Generate response
messages = [
    {"role": "system", "content": "You are a helpful AI assistant."},
    {"role": "user", "content": "Explain quantum computing in simple terms."}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(
    inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

🎓 Training Data

The model was trained on 10,887 diverse conversations from ShareGPT52K, covering:

💻 Technical assistance and programming
📚 General knowledge and education
✍️ Creative writing and storytelling
🔧 Problem-solving and troubleshooting
💬 Natural conversation and dialogue
🌍 Wide range of topics and domains

✨ What Makes This Model Special

High-Quality Data: Trained exclusively on ShareGPT conversations, known for natural and helpful responses
Optimized for Voice AI: Designed for conversational applications with natural dialogue flow
Efficient LoRA: Only 161MB adapter that works with the base model
Well-Balanced: 3 full epochs ensure good learning without overfitting
Conversational Excellence: Excels at multi-turn dialogues and context understanding

💪 Model Strengths

Natural, human-like conversational responses
Strong context retention across multiple turns
Helpful and informative explanations
Creative problem-solving approaches
Adaptable to various conversation styles

⚠️ Limitations

Requires base Llama 3 8B model (this is just the LoRA adapter)
Trained primarily on English conversations
May inherit biases present in training data
Requires proper chat template formatting
Best performance with conversational use cases

📈 Performance

The model shows improved performance over the base model on:

Conversational coherence
Response helpfulness
Natural dialogue flow
Context understanding
Multi-turn conversations

🛠️ Use Cases

Perfect for:

🎙️ Voice assistants
💬 Chatbots
📞 Customer service AI
🤖 Interactive AI applications
📱 Mobile AI assistants
🌐 Web-based chat interfaces

📦 Model Files

adapter_model.safetensors - LoRA adapter weights (~161MB)
adapter_config.json - Adapter configuration
tokenizer_config.json - Tokenizer settings
special_tokens_map.json - Special tokens
Checkpoints saved every 100 steps

🔧 Hardware Requirements

For Inference:

GPU: 12GB+ VRAM (with 4-bit quantization)
RAM: 16GB+ system memory
Storage: ~5GB (base model) + 161MB (adapter)

Recommended Setup:

GPU: RTX 3090, RTX 4090, A100, or similar
With 4-bit quantization: Can run on consumer GPUs

📝 Citation

bibtex

@misc{llama3-sharegpt-10k-voice-ai,
  author = {owenergy},
  title = {Llama 3 8B ShareGPT 10K Voice AI},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/owenergy/llama3-sharegpt-10k-voice-ai},
  note = {LoRA finetuned on 10,887 ShareGPT conversations}
}

📄 License

This model inherits the Llama 3 license from Meta. Please review the Llama 3 License Agreement.

🙏 Acknowledgments

Meta for the Llama 3 base model
RyokoAI for the ShareGPT52K dataset
HuggingFace for the transformers and PEFT libraries

📧 Contact

For questions, issues, or feedback, please open an issue on HuggingFace.

Model Card by: owenergy
Date: December 2025
Status: Production Ready ✅