abhishekgahlot/better-sql-agent-llama

Better SQL Agent - Llama 3.1 8B

Training Results

• Training Samples: 19,480 (SQL analytics + technical conversations)
• Hardware: NVIDIA 4x A10G GPU (96GB VRAM)

Model Description

This is a high-performance fine-tuned version of Meta-Llama-3.1-8B-Instruct, specifically optimized for:

• SQL query generation and optimization
• Data analysis and insights
• Technical assistance and debugging
• Tool-based workflows

Training Configuration

• Base Model: meta-llama/Llama-3.1-8B-Instruct
• Training Method: LoRA (Low-Rank Adaptation)
  • Rank: 16, Alpha: 32, Dropout: 0.05
• Quantization: 4-bit with BF16 training precision
• Context Length: 128K tokens (extended from base)
• Optimizer: AdamW with cosine scheduling

Quick Start

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load the fine-tuned model
model_name = "abhishekgahlot/better-sql-agent-llama"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)

# Generate SQL query
prompt = """<|begin_of_text|><|start_header_id|>user<|end_header_id|>

Create a SQL query to find the top 5 customers by total revenue in 2024:

<|eot_id|><|start_header_id|>assistant<|end_header_id|>

"""

inputs = tokenizer(prompt, return_tensors="pt")
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.7,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], 
                          skip_special_tokens=True)
print(response)

Performance Metrics

Metric	Value
Starting Loss	1.53
Final Loss	0.0508
Loss Reduction	96.7%
Training Time	8.9 hours

Use Cases

• SQL Generation: Create complex queries from natural language
• Data Analysis: Generate insights and analytical queries
• Code Assistance: Debug and optimize SQL code
• Technical Support: Answer database and analytics questions
• Learning Aid: Explain SQL concepts and best practices

Training Data

The model was trained on a curated dataset of 19,480 high-quality examples including:

• SQL query generation tasks
• Data analysis conversations
• Technical problem-solving dialogues
• Tool usage patterns and workflows

Optimization Features

• 4-bit Quantization: Reduced memory footprint
• Flash Attention: Optimized attention mechanism
• Mixed Precision: BF16 training for efficiency

License

This model inherits the Llama 3.1 license from the base model. Please review the official license for usage terms.

Acknowledgments

• Based on Meta's Llama 3.1 8B Instruct model

Model Card Contact

For questions about this model, please open an issue in the repository or contact the model author.

---