| --- |
| license: gemma |
| base_model: unsloth/gemma-3-4b-it |
| tags: |
| - gguf |
| - gemma |
| - finetuned |
| - affiliate |
| - screening |
| - conversational |
| - ollama |
| - llama-cpp |
| - unsloth |
| - q4_k_m |
| language: |
| - en |
| - bn |
| pipeline_tag: text-generation |
| --- |
| |
| # 🤖 AI Manager - Affiliate Screening Assistant |
|
|
| ## 📌 Overview |
|
|
| **AI Manager** is a fine-tuned version of Google's **Gemma-3-4B** model, specifically trained for **affiliate application screening**. It acts as an intelligent assistant that verifies affiliate applications, enforces network policies, and communicates in **English, Bangla, or Banglish** based on user preference. |
|
|
| This model is designed for **Monir Hasan (Monir iMonetizeIt)**, Sales Manager & Affiliate Specialist at **iMonetizeIt**, a global CPA and Smartlink affiliate network. |
|
|
| --- |
|
|
| ## 🎯 Use Cases |
|
|
| - ✅ **Affiliate Application Screening** – Verify applicant eligibility |
| - ✅ **Policy Enforcement** – Ensure compliance with network rules |
| - ✅ **Document Verification** – Check NID, screenshots, and other documents |
| - ✅ **Multi-language Support** – Respond in English, Bangla, or Banglish |
| - ✅ **Rule-based Decision Making** – Enforce 18+ age limit, desktop screenshots, etc. |
|
|
| --- |
|
|
| ## 🛠️ Training Details |
|
|
| | Parameter | Value | |
| |-----------|-------| |
| | **Base Model** | `unsloth/gemma-3-4b-it` | |
| | **Fine-tuning Method** | QLoRA (4-bit quantization) | |
| | **LoRA Rank (r)** | 32 | |
| | **LoRA Alpha** | 64 | |
| | **Trainable Parameters** | 65.5M (1.5% of total) | |
| | **Dataset Size** | 584 training examples | |
| | **Evaluation Size** | 64 validation examples | |
| | **Epochs** | 4 | |
| | **Batch Size** | 2 | |
| | **Learning Rate** | 2e-4 | |
| | **Optimizer** | AdamW 8-bit | |
| | **Loss Function** | Cross-entropy with response-only masking | |
| | **Hardware** | Kaggle T4 GPU (2x) | |
|
|
| ### Training Progress |
|
|
| | Epoch | Training Loss | Validation Loss | |
| |-------|---------------|-----------------| |
| | 1 | 2.439 | 1.902 | |
| | 2 | 0.709 | 1.683 | |
| | 3 | 0.453 | 1.695 | |
| | 4 | 0.385 | 1.702 | |
|
|
| --- |
|
|
| ## 📁 Model Files |
|
|
| | File | Size | Description | |
| |------|------|-------------| |
| | `gemma-3-4b-it.Q4_K_M.gguf` | 2.49 GB | Quantized GGUF model (Q4_K_M) | |
| | `gemma-3-4b-it.F16-mmproj.gguf` | 812 MB | Multimodal projection file | |
|
|
| --- |
|
|
| ## 🚀 How to Use |
|
|
| ### Option 1: Ollama (Recommended) |
|
|
| Create a `Modelfile`: |
|
|
| ```dockerfile |
| FROM https://huggingface.co/imonetizeitbd/ai-manager/resolve/main/gemma-3-4b-it.Q4_K_M.gguf |
| |
| TEMPLATE """<bos><start_of_turn>user |
| {{ .Prompt }}<end_of_turn> |
| <start_of_turn>model |
| {{ .Response }}<end_of_turn>""" |
| |
| PARAMETER temperature 0.3 |
| PARAMETER top_p 0.9 |
| PARAMETER stop "<end_of_turn>" |
