Sri-Vigneshwar-DJ/hawky-ai-Qwen3-0.6B-Marketing-MoT

Qwen3-0.6B - Performance Marketing MoT (Hawky AI)

This model is a fine-tuned version of Qwen3-0.6B, specialized in Performance Marketing Critical Thinking. It utilizes the Mixture of Thoughts (MoT) reasoning framework to provide deep, multi-perspective strategic analysis.

Model Description

• Developed by: Sri-Vigneshwar-DJ (Hawky AI)
• Model Type: Causal Language Model (Fine-tuned for Reasoning)
• Base Model: Qwen/Qwen3-0.6B
• Language(s): English
• License: Based on Qwen3 license
• Training Data: 1,500 records of high-quality Performance Marketing Critical Thinking scenarios.

Mixture of Thoughts (MoT) Framework

The model is trained to follow 5 distinct cognitive reasoning styles, allowing it to handle complex marketing challenges with depth and rigour:

1. Linear Reasoning (Sequential Logic): Focuses on audit-to-execution workflows.
2. Exploratory Reasoning (Strategic Paths): Evaluates multiple approaches (Aggressive/Conservative/Innovative) before selecting the best one.
3. Critical & Reflective Reasoning (Self-Correction): Drafts a plan, critiques its own internal logic for risks, and provides a robust final output.
4. Deconstructive Reasoning (First Principles): Breaks down complex goals into basic marketing axioms (CAC, LTV, CPM, CVR).
5. Analogical Reasoning (Expert Perspctive): Views problems through specialized lenses (e.g., Risk Manager, Investment Banker) to identify non-obvious constraints.

Intended Uses & Limitations

Intended Uses

• Strategic planning for Meta Ads, Google Ads, and TikTok.
• Troubleshooting attribution discrepancies (GA4 vs Ad Platforms).
• CRO and Landing Page optimization audits.
• Unit economics and LTV/CAC analysis.
• Creative strategy and fatigue management.

Limitations

• The model is small (0.6B parameters), and while highly efficient, it may occasionally over-simplify extremely complex cross-channel financial modeling.
• Knowledge cutoff is based on the 2024-2025 marketing landscape.

Training Details

• Method: QLoRA (4-bit quantization + LoRA adapters)
• Optimizer: AdamW 8-bit
• Batch Size: 2 (Gradient Accumulation: 4)
• Learning Rate: 2e-4
• Precision: BF16/FP16 (depending on hardware)
• Training Platform: Google Colab (Tesla T4) via Unsloth

How to Use

The model follows the Qwen ChatML-like format. For the best reasoning results, use a system prompt that specifies an MoT style.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Sri-Vigneshwar-DJ/hawky-ai-Qwen3-0.6B-Marketing-MoT"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

# Example input with an exploratory MoT prompt
system_msg = "You are a Performance Marketing expert. Identify at least three possible approaches to this marketing problem. Evaluate the pros and cons of each based on performance data, then select the best one."
user_msg = "Our Meta ads CPA has risen by 40% in the last 7 days despite stable ROAS. What should we investigate?"

prompt = f"<|im_start|>system\n{system_msg}<|im_end|>\n<|im_start|>user\n{user_msg}<|im_end|>\n<|im_start|>assistant\n<thinking>"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Citation & Acknowledgements

Special thanks to the Qwen team for the base model and Unsloth for the optimization tools.