Downloads · 30 days
0
Sri-Vigneshwar-DJ/hawky-ai-h1-mini-4b-experimental
hawky-ai-h1-mini-4b-experimental is a text generation model from Sri-Vigneshwar-DJ. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as gemma.
<p align="center" <a href="https://hawky.ai"<img src="https://img.shields.io/badge/🦅Hawky.ai-VisitWebsite-blue?style=for-the-badge" alt="Hawky.ai Website"/</a <img src="https://img.shields.io/badge/Base-Gemma%203%204…
Downloads · 30 days
0
Access
Public
Updated Jan 8, 2026
Repo size
169 MB
Likes
1
Public
Click a slice to open those files.
.safetensors131 MB · 77%
From the Hugging Face model README
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.
| 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 |
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.
┌─────────────────────────────────────────────────────────────────────────────┐
│ 🦅 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 │
└─────────────────────────────────────────────────────────────────────────────┘
| Property | Value |
|---|---|
| Model | Google Gemma 3 4B Instruct |
| Parameters | 4 Billion |
| Architecture | Decoder-only Transformer |
| Context Length | 128K tokens |
| 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 |
| 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 |
| 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 |
| Metric | Value |
|---|---|
| Final Training Loss | TBD |
| Training Hardware | NVIDIA Tesla T4 (Google Colab) |
| Training Duration | ~1-1.5 hours |
pip install transformers peft accelerate bitsandbytes torch
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")
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)
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.
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.
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.
| 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 |
| 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 |
✅ Recommended Uses:
❌ Not Recommended For:
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.
@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}}
}
This model is released under the Gemma License.
The fine-tuning code and dataset methodology are released under Apache 2.0.