CAI-1-32B: Marketing AI with Tool-Calling

A fine-tuned version of Qwen3-32B specialized for marketing tasks with native tool-calling capabilities.

(This model is a prototype and not designed for user interaction its meant to serve as an Neural RAG the first of it's kind via a MCP server)

Model Overview

CAI-1-32B is a 32B parameter model fine-tuned for marketing expertise with native tool-calling capabilities.

Training Details

• Base Model: Qwen3-32B
• Fine-tuning Method: QLoRA (4-bit quantization + LoRA adapters)
• Training Hardware: H200 GPU (143GB VRAM)
• Training Duration: 7 hours
• Training Data:
  • Batch 1: Marketing knowledge and strategies
  • Batch 2: 3,000 examples (33.1% with tool-calling)
• Tool Functions: 7 marketing-specific tools including:
  • optimize_email_sequence (215 examples)
  • generate_sales_copy (183 examples)
  • analyze_creative_performance (134 examples)
  • And more...

Performance

• Evaluation Loss: 0.222 (57.5% reduction from baseline)
• Generalization: Excellent (no overfitting detected)
• Grade: A+ (98.1/100 in comprehensive evaluation)

Optimal Usage

1. For Tool-Calling

Temperature: Use 0.1-0.3 for consistent JSON formatting

# Example prompt that triggers tool usage
prompt = "Can you optimize our email sequence for better conversions?"

# Generation settings
generation_config = {
    "temperature": 0.3,  # Low for tool consistency
    "max_new_tokens": 500,
    "do_sample": True,
    "top_p": 0.9
}

Trigger Words that activate tool usage:

• "optimize" (4.3% of tool examples)
• "analyze" (5.3% of tool examples)
• "automate" (1.8% of tool examples)
• "campaign", "segment", "personalize"

Expected Response Pattern:

Assistant: I'll optimize email sequence for conversions to provide detailed insights.
{
  "tool_calls": [{
    "id": "call_xxx",
    "type": "function",
    "function": {
      "name": "optimize_email_sequence",
      "arguments": "{...}"
    }
  }]
}
[Detailed marketing strategy follows...]

2. For Marketing Expertise

Temperature: Use 0.5-0.7 for creative marketing content

# Example for marketing strategy
prompt = "What's the best strategy to increase customer retention for a SaaS product?"

generation_config = {
    "temperature": 0.7,  # Higher for creativity
    "max_new_tokens": 800,
    "do_sample": True,
    "top_p": 0.9
}

Key Insights

1. Tool-Calling Pattern: The model ALWAYS explains what it's doing before calling a tool ("I'll [action] to provide detailed insights")

2. Marketing Domain: Specialized in:

   • Email marketing optimization
   • Sales copy generation
   • Creative performance analysis
   • A/B testing frameworks
   • Statistical significance calculations

3. Best Practices:

   • Use action-oriented prompts for tool activation
   • Lower temperature (0.1-0.3) for tool JSON consistency
   • Higher temperature (0.5-0.7) for creative marketing content
   • Include marketing context in system prompts

Example Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model
model = AutoModelForCausalLM.from_pretrained(
    "tigres2526/CAI-1-32B",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    load_in_4bit=True  # Optional: for memory efficiency
)
tokenizer = AutoTokenizer.from_pretrained("tigres2526/CAI-1-32B")

# Tool-calling example
messages = [
    {"role": "system", "content": "You are a marketing AI with access to optimization tools."},
    {"role": "user", "content": "Analyze our email campaign performance and suggest improvements"}
]

# Apply chat template
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")

# Generate with low temperature for tools
outputs = model.generate(
    **inputs,
    max_new_tokens=500,
    temperature=0.3,
    do_sample=True,
    top_p=0.9
)

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

Troubleshooting

Issue: Model doesn't output tool calls

• Solution: Use trigger words + lower temperature (0.1-0.3)
• Check: Ensure prompts contain action verbs (optimize, analyze, etc.)

Issue: Inconsistent JSON formatting

• Solution: Set temperature to 0.1 or use greedy decoding (do_sample=False)

Issue: Generic responses instead of marketing expertise

• Solution: Include marketing context in system prompt
• Example: "You are a senior marketing strategist with 20 years of experience..."

Training Details

• Learning Rate: 1e-5
• Batch Size: 16 (effective)
• LoRA Config: r=64, alpha=256
• Gradient Checkpointing: Enabled
• Dynamic Padding: Enabled

License

Same as Qwen2.5-32B base model.

Citation

If you use this model, please cite:

@misc{cai1-32b,
  title={CAI-1-32B: Marketing AI with Tool-Calling},
  author={tigres2526},
  year={2024}
}