🦅 Hawky AI H1 Mini 4B (Experimental)

Built by Hawky.ai — AI-Powered Creative Intelligence for Performance Marketing

📖 Model Description

Hawky AI H1 Mini 4B Experimental is a domain-specific fine-tuned version of Google's Gemma 3 4B Instruct (4 Billion parameters), trained specifically for Meta Andromeda algorithm expertise and creative-first performance marketing strategy.

This experimental model was created by Hawky.ai to measure the effectiveness of domain-specific fine-tuning for performance marketing knowledge, particularly around Meta's latest advertising AI systems.

🎯 What This Model Knows

Domain	Capabilities
🎯 Meta Andromeda Algorithm	Deep understanding of Meta's AI-powered ad retrieval system, how it matches creatives to users, and optimization strategies
🎨 Creative Strategy	Best practices for ad creative diversity, video hooks, creative fatigue, and format selection
📊 Campaign Structure	Guidance on Advantage+ Shopping, CBO vs ABO, budget allocation, and campaign consolidation
📈 Performance Metrics	Interpretation of CPM, CPA, ROAS signals and diagnostic frameworks
👥 Audience Strategy	Broad vs narrow targeting under Andromeda, audience signals, and creative-as-targeting

🧠 Chain-of-Thought Reasoning

The model was trained with Chain-of-Thought (CoT) reasoning, meaning it thinks through problems step-by-step before providing answers. Responses include <think> blocks showing the reasoning process.

🧠 Performance Marketing Critical Thinking Framework

┌─────────────────────────────────────────────────────────────────────────────┐
│                    🦅 HAWKY AI H1 MINI 4B                                    │
│              Performance Marketing Critical Thinking Framework               │
└─────────────────────────────────────────────────────────────────────────────┘

                              ┌─────────────────┐
                              │   📥 USER QUERY  │
                              │  Marketing Q&A   │
                              └────────┬────────┘
                                       │
                                       ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                        🧠 <think> REASONING PHASE                            │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│   ┌──────────────┐     ┌──────────────┐     ┌──────────────┐               │
│   │ 1️⃣ CLASSIFY   │────▶│ 2️⃣ CONTEXT   │────▶│ 3️⃣ ANALYZE   │               │
│   │              │     │              │     │              │               │
│   │ • Domain ID  │     │ • Andromeda  │     │ • Framework  │               │
│   │ • Query Type │     │ • Best Prac. │     │ • Trade-offs │               │
│   └──────────────┘     └──────────────┘     └──────────────┘               │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘
                                       │
                                       ▼
┌─────────────────────────────────────────────────────────────────────────────┐
│                         📚 KNOWLEDGE DOMAINS                                 │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                             │
│  ┌───────────┐ ┌───────────┐ ┌───────────┐ ┌───────────┐ ┌───────────┐     │
│  │    🎯     │ │    🎨     │ │    📊     │ │    📈     │ │    👥     │     │
│  │ ANDROMEDA │ │ CREATIVE  │ │ CAMPAIGN  │ │PERFORMANCE│ │ AUDIENCE  │     │
│  │ ALGORITHM │ │ STRATEGY  │ │ STRUCTURE │ │  METRICS  │ │ STRATEGY  │     │
│  ├───────────┤ ├───────────┤ ├───────────┤ ├───────────┤ ├───────────┤     │
│  │• AI Match │ │• 8-15 Ads │ │• Adv+/CBO │ │• CPM/CPA  │ │• Broad    │     │
│  │• Signals  │ │• Hooks    │ │• Budget   │ │• ROAS     │ │• Signals  │     │
│  │• Retrieval│ │• Formats  │ │• Testing  │ │• Diagnose │ │• Creative │     │
│  └───────────┘ └───────────┘ └───────────┘ └───────────┘ └───────────┘     │
│                                                                             │
└─────────────────────────────────────────────────────────────────────────────┘
                                       │
                                       ▼
                         ┌─────────────────────────┐
                         │    ⚡ SYNTHESIZE         │
                         │  Combine & Prioritize   │
                         └────────────┬────────────┘
                                      │
                    ┌─────────────────┼─────────────────┐
                    ▼                 ▼                 ▼
           ┌──────────────┐  ┌──────────────┐  ┌──────────────┐
           │📋 STRATEGIC  │  │💡 REASONING  │  │⚡ NEXT STEPS │
           │RECOMMENDATION│  │ EXPLANATION  │  │ & QUICK WINS │
           └──────────────┘  └──────────────┘  └──────────────┘

┌─────────────────────────────────────────────────────────────────────────────┐
│  🦅 Hawky.ai — AI-Powered Creative Intelligence │ https://hawky.ai          │
└─────────────────────────────────────────────────────────────────────────────┘

📊 Training Details

Base Model

Property	Value
Model	Google Gemma 3 4B Instruct
Parameters	4 Billion
Architecture	Decoder-only Transformer
Context Length	128K tokens

Fine-tuning Configuration

Property	Value
Method	QLoRA (4-bit Quantization + LoRA)
LoRA Rank (r)	16
LoRA Alpha	32
LoRA Dropout	0.05
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Trainable Parameters	~0.5% of total

Training Hyperparameters

Property	Value
Epochs	3
Effective Batch Size	16
Learning Rate	2e-4
LR Scheduler	Cosine with warmup
Warmup Ratio	0.1
Precision	FP16
Max Sequence Length	1024
Optimizer	Paged AdamW 8-bit

Training Data

Property	Value
Dataset	Hawky Andromeda Dataset
Samples	1,000 Q&A pairs
Format	Chain-of-Thought with `<think>` reasoning blocks
Topics	Meta Andromeda, creative strategy, campaign structure, performance optimization

Training Results

Metric	Value
Final Training Loss	TBD
Training Hardware	NVIDIA Tesla T4 (Google Colab)
Training Duration	~1-1.5 hours

🚀 Quick Start

Installation

pip install transformers peft accelerate bitsandbytes torch

Load the Model

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

# Quantization config (optional, for memory efficiency)
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "google/gemma-3-4b-it",
    quantization_config=bnb_config,
    device_map="auto",
    torch_dtype=torch.float16,
)

# Load fine-tuned LoRA adapter
model = PeftModel.from_pretrained(
    base_model,
    "Ashwath21/hawky-ai-h1-mini-4b-experimental"
)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("Ashwath21/hawky-ai-h1-mini-4b-experimental")

Generate Responses

def ask_hawky(question):
    """Ask the Hawky AI model a performance marketing question."""
    
    # Gemma 3 chat format
    prompt = f"<start_of_turn>user\n{question}<end_of_turn>\n<start_of_turn>model\n"
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=512,
            temperature=0.7,
            top_p=0.9,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id,
        )
    
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return response.split("<start_of_turn>model")[-1].strip()

# Example
response = ask_hawky("How many creatives should I use in my Advantage+ campaign?")
print(response)

💬 Example Conversations

Creative Diversity

Q: How many distinct creative concepts should I use in my Advantage+ campaign?

<think>
This is a creative strategy question about Andromeda's diversity requirements.

Key considerations:
- Andromeda treats similar variations as one ad
- Genuine diversity means different hooks, angles, formats
- The algorithm needs variety to optimize effectively

Research indicates 8-15 genuinely distinct concepts is optimal.
</think>

You should aim for 8-15 genuinely distinct creative concepts per campaign.
"Genuinely distinct" means different hooks, angles, formats, or visual styles -
not just minor copy variations. Five product shots with slightly different text
are treated as ONE ad by Andromeda. True diversity gives the algorithm more
tools to match the right creative to each user.

Video Hooks

Q: What makes an effective video hook under Andromeda?

<think>
This is about the critical first 1-3 seconds of video ads.

Under Andromeda, hooks serve two purposes:
1. Stop the scroll (user engagement)
2. Signal audience relevance (algorithm matching)

Effective hooks share certain characteristics.
</think>

Effective video hooks capture attention in the first 1-3 seconds through:
- Pattern interrupts (unexpected visuals)
- Direct benefits ("This saved me $800")
- Compelling questions
- Social proof ("500,000 customers")
- Before/after reveals

Design for sound-off with text overlays. Test 3-5 different hooks per core
video concept - hooks are the highest-impact creative element and fatigue
fastest, so refresh weekly.

Performance Diagnosis

Q: My CPM is rising but conversions are stable. What does this indicate?

<think>
Let me analyze this performance pattern.

CPM rising + conversions stable means:
- Cost per impression is increasing
- But conversion volume hasn't dropped

This is actually a positive signal under Andromeda.
</think>

This is actually a healthy signal! Rising CPM with stable conversions indicates
Andromeda is finding higher-value users who cost more to reach but convert at
better rates. Your effective CPA is likely improving.

Monitor your CPA and ROAS - if those metrics are stable or improving, don't
intervene. The algorithm is optimizing for outcomes, not reach efficiency.

🔄 Hawky AI Model Family

Model	Base	Parameters	Status	Use Case
H1 Mini 1B	Gemma 3 1B	1B	✅ Released	Fast inference, edge deployment
H1 Mini 4B (this)	Gemma 3 4B	4B	✅ Released	Better reasoning, balanced performance
H1 8B	Gemma 3 8B	8B	🔜 Planned	Enhanced capability
H1 Pro	Gemma 3 27B	27B	🔜 Planned	Full capability, complex analysis

⚠️ Limitations

Limitation	Description
Experimental	This is a research model, not production-ready
Domain-Specific	Optimized for Meta advertising; may not perform well on general tasks
Knowledge Cutoff	Training data reflects best practices as of January 2026
Model Size	4B parameters - good balance of speed and capability
No Real-Time Data	Cannot access current campaign data or live metrics

🎯 Intended Use Cases

✅ Recommended Uses:

Research on domain-specific fine-tuning effectiveness
Learning about Meta Andromeda and creative strategy
Prototyping AI-assisted marketing tools
Educational demonstrations
Brainstorming creative strategies

❌ Not Recommended For:

Production deployment without further evaluation
Financial or business-critical decisions
Replacing human marketing expertise
Real-time campaign management

🏢 About Hawky.ai

🦅 Visit Hawky.ai

Hawky.ai is an AI-powered creative intelligence platform for digital marketing, providing:

Feature	Description
🔍 Automated Campaign Analysis	Deep insights into ad performance across platforms
🎨 Creative Performance Optimization	Data-driven creative recommendations and scoring
📊 Competitive Intelligence	Track and analyze competitor ad strategies
🤖 AI-Powered Insights	Actionable recommendations for performance marketers

Trusted by: WPP, Madison, GroupM, TVS Motors, Tanishq, Bajaj Finserv, and leading brands & agencies.

👉 Learn more at hawky.ai

📚 Citation

@misc{hawky-ai-h1-mini-4b-2026,
  author = {Hawky.ai},
  title = {Hawky AI H1 Mini 4B Experimental: A Gemma 3 4B Fine-tuned Model for Performance Marketing},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Ashwath21/hawky-ai-h1-mini-4b-experimental}}
}

📄 License

This model is released under the Gemma License.

The fine-tuning code and dataset methodology are released under Apache 2.0.

🙏 Acknowledgments

Google for the Gemma 3 4B base model
Hugging Face for the transformers and PEFT libraries
Meta for publicly sharing Andromeda algorithm insights

Built with ❤️ by Hawky.ai
https://hawky.ai

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for Sri-Vigneshwar-DJ/hawky-ai-h1-mini-4b-experimental

Base model

google/gemma-3-4b-pt

Finetuned

google/gemma-3-4b-it

Adapter

(254)

this model