# Hawky-AI H1 4B Performance Marketing (hawky-ai-H1-4b-PM)

<div align="center">



**The first open-source LLM fine-tuned specifically for Performance Marketing expertise**

[![Model on HF](https://huggingface.co/datasets/huggingface/badges/resolve/main/model-on-hf-md.svg)](https://huggingface.co/Sri-Vigneshwar-DJ/hawky-ai-H1-4b-PM)
[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
[![Built by Hawky.ai](https://img.shields.io/badge/Built%20by-Hawky.ai-orange)](https://hawky.ai)

</div>

## Model Description

**Hawky-AI H1 4B PM** is a specialized language model fine-tuned for performance marketing tasks. Built on Google's Gemma 3 4B architecture, this model was trained using knowledge distillation from Claude Opus 4.5, capturing expert-level reasoning for paid media optimization, creative strategy, and campaign management.

This model is developed by [Hawky.ai](https://hawky.ai), the Creative Intelligence Platform for Performance Marketing, serving major agencies (WPP, Madison, GroupM) and brands (TVS Motors, Tanishq, Bajaj Finserv).

### Key Features

- 🎯 **Domain-Specialized**: Purpose-built for performance marketing, not a general-purpose model
- 🧠 **Expert Reasoning**: Distilled from Claude Opus 4.5 with chain-of-thought marketing expertise
- ⚡ **Efficient**: 4B parameters - runs on consumer GPUs (8GB+ VRAM)
- 📊 **Practical**: Trained on real-world scenarios from Meta Ads, Google Ads, TikTok, and more
- 🔓 **Open Source**: Fully open weights for the marketing community

## Intended Use

### Primary Use Cases

| Use Case | Description |
|----------|-------------|
| **Campaign Troubleshooting** | Diagnose ROAS drops, CTR declines, high CPAs |
| **Strategy Recommendations** | Campaign structure, budget allocation, scaling strategies |
| **Creative Analysis** | Hook rate optimization, fatigue detection, A/B testing |
| **Platform Expertise** | Meta Ads, Google Ads, TikTok, Performance Max guidance |
| **Measurement & Attribution** | ROAS vs MER analysis, incrementality, LTV optimization |

### Target Users

- Performance Marketers
- Media Buyers
- Growth Teams
- Marketing Agencies
- D2C Brand Teams

## Training Details

### Base Model
- **Architecture**: Gemma 3 4B Instruct
- **Parameters**: 4 Billion
- **Context Length**: 8,192 tokens

### Fine-tuning Approach
- **Method**: QLoRA (4-bit quantization with LoRA adapters)
- **Teacher Model**: Claude Opus 4.5 (Anthropic)
- **Technique**: Knowledge Distillation with Chain-of-Thought reasoning
- **Training Data**: Curated performance marketing scenarios and expert responses

### Training Configuration

```yaml
LoRA Config:
  r: 64
  lora_alpha: 128
  target_modules: [q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]
  lora_dropout: 0.05

Training Args:
  epochs: 3
  learning_rate: 2e-4
  batch_size: 2 (effective: 16 with gradient accumulation)
  optimizer: paged_adamw_8bit
  scheduler: cosine
  precision: bf16
```

### Training Data Domains

| Domain | Topics Covered |
|--------|---------------|
| **Meta Ads** | Campaign structure, ASC vs manual, bidding strategies, retargeting, scaling, creative fatigue |
| **Google Ads** | Quality Score, Performance Max, lead gen, Search optimization |
| **Creative Strategy** | Hook rates, A/B testing, funnel-stage creative, TikTok native |
| **Measurement** | Attribution (ROAS/MER), incrementality testing, LTV:CAC, UTM tracking |
| **Strategy** | Budget allocation, competitive intelligence, landing page optimization |

## Usage

### Quick Start with Transformers

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

model_id = "Sri-Vigneshwar-DJ/hawky-ai-H1-4b-PM"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Example: Diagnose a campaign issue
prompt = """<bos><start_of_turn>user
My Meta ads CTR dropped from 2.1% to 0.8% over two weeks. Frequency is at 4.5. What's happening and what should I do?<end_of_turn>
<start_of_turn>model
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=1024,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)

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

### With 4-bit Quantization (Low VRAM)

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

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True
)

model = AutoModelForCausalLM.from_pretrained(
    "Sri-Vigneshwar-DJ/hawky-ai-H1-4b-PM",
    quantization_config=bnb_config,
    device_map="auto"
)
```

### Example Prompts

```python
# Campaign Troubleshooting
"A D2C brand's ROAS dropped from 3.5x to 1.8x over a month. They're running TOF broad, MOF retargeting, and BOF cart abandonment campaigns at $2000/day. Frequency is 2.1, same creatives for 6 weeks. Diagnose and provide an action plan."

# Strategy Question
"Should I use Advantage+ Shopping Campaign or manual campaigns for a new e-commerce brand with limited pixel data?"

# Creative Analysis
"Explain hook rate optimization for video ads. How do I diagnose and fix poor hook rates?"

# Measurement
"What's the difference between ROAS, MER, and blended metrics? When should I use each?"

# Scaling
"What's the best way to scale a Meta campaign from $500/day to $5000/day without killing performance?"
```

## Evaluation

### Qualitative Assessment

The model was evaluated on held-out performance marketing scenarios:

| Capability | Assessment |
|------------|------------|
| Platform Mechanics Accuracy | ✅ Strong - Correct Meta/Google feature knowledge |
| Strategic Reasoning | ✅ Strong - Logical diagnostic frameworks |
| Actionable Recommendations | ✅ Strong - Specific, implementable advice |
| Chain-of-Thought Quality | ✅ Strong - Clear step-by-step reasoning |
| Edge Case Handling | ⚡ Good - Handles most scenarios well |

### Comparison to Base Model

| Aspect | Base Gemma 3 4B | Hawky-AI H1 4B PM |
|--------|-----------------|-------------------|
| Marketing terminology | Generic | Domain-specific |
| Platform mechanics | Surface-level | Expert-level detail |
| Diagnostic frameworks | None | Structured approaches |
| Recommendations | Generic advice | Specific, actionable |

## Limitations

- **Knowledge Cutoff**: Training data reflects marketing best practices as of early 2025. Platform features may have changed.
- **Platform Specifics**: Strongest on Meta and Google Ads; other platforms have less coverage.
- **No Real-Time Data**: Cannot access live campaign data or current market conditions.
- **Not Financial Advice**: Recommendations are educational; always validate with your own testing.
- **English Only**: Optimized for English language queries.

## Ethical Considerations

This model is designed to assist performance marketers with strategic and tactical decisions. Users should:

- Validate recommendations against platform documentation
- Test strategies at small scale before full implementation
- Consider privacy and data protection when implementing targeting strategies
- Follow platform advertising policies and guidelines

## Citation

```bibtex
@misc{hawky-ai-h1-4b-pm,
  author = {Sri Vigneshwar DJ and Hawky.ai Team},
  title = {Hawky-AI H1 4B Performance Marketing: A Domain-Specialized LLM for Paid Media},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Sri-Vigneshwar-DJ/hawky-ai-H1-4b-PM}}
}
```

## About Hawky.ai

[Hawky.ai](https://hawky.ai) is a Creative Intelligence Platform for Performance Marketing, helping brands and agencies transform creative guesswork into data-driven decisions. Our platform provides:

- **Creative Analyzer**: AI pattern recognition using performance data
- **Competitor 360**: Competitive intelligence and strategy analysis  
- **Trend Analyzer**: Emerging signal tracking and winning trend prediction

We serve major agencies (WPP, Madison, GroupM) and brands (TVS Motors, Tanishq, Bajaj Finserv) across India and beyond.

## Links

- 🌐 **Website**: [hawky.ai](https://hawky.ai)
- 🤗 **Hugging Face**: [Hawky-ai](https://huggingface.co/Hawky-ai)
- 💼 **LinkedIn**: [Hawky.ai](https://linkedin.com/company/hawky-ai)
- 📧 **Contact**: team@hawky.ai

## Acknowledgments

- **Anthropic** for Claude Opus 4.5 used as the teacher model
- **Google** for the Gemma 3 base model architecture
- **Hugging Face** for the transformers and PEFT libraries
- The open-source ML community

---

<div align="center">

**Built with ❤️ by [Hawky.ai](https://hawky.ai) - Empowering marketers with AI-driven clarity**

</div>