HIBA-7B-Soul

<div align="center"> <br> <h1>🌸 HIBA-7B-Soul (هبة)</h1> <h3>The World's First High-Fidelity Therapeutic AI Specialist</h3> <p> <img src="https://img.shields.io/badge/License-Apache%202.0-blue?style=for-the-badge" alt="License"/> <img src="https://img.shields.io/badge/Version-1.0.0-orange?style=for-the-badge" alt="Version"/> <img src="https://img.shields.io/badge/Model%20Size-7B%20Params-green?style=for-the-badge" alt="Size"/> <a href="https://github.com/boubli/HIBA"><img src="https://img.shields.io/badge/GitHub-Repo-teal?style=for-the-badge&logo=github" alt="GitHub"/></a> </p> <p> <a href="https://colab.research.google.com/github/boubli/HIBA/blob/main/HIBA_Quickstart.ipynb" target="_blank"> <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/> </a> </p> </div>

---

🔬 Executive Summary: The Specialist vs. The Generalists

In the era of GPT-5 and Llama 4 (405B), why does the world need a 7B model?

Because scale ignores the soul.

While generalist models maximize MMLU scores, HIBA-7B-Soul maximizes Human-Centric Empathy (HCE). Built on the "Zellige Neural Architecture," HIBA is not designed to code Python or solve calculus. It is designed for one purpose: to sit with you in the dark until you find the light.

---

🧠 Neural Architecture & Design

HIBA modifies the standard Transformer architecture by injecting specialized "Soul Adapters" into the attention mechanism.

graph TD
    A[User Input] --> B[Qwen 2.5 Tokenizer];
    B --> C{Soul Gate};
    C -- General query --> D[Frozen Qwen Blocks];
    C -- Emotional query --> E[LoRA Adapters];
    E --> F[Cultural Context Layer];
    F --> G[Empathy Refinement Head];
    D --> H[Output Generation];
    G --> H;
    H --> I[Final Response];
    
    style E fill:#f9f,stroke:#333,stroke-width:2px
    style F fill:#bbf,stroke:#333,stroke-width:2px

---

📊 Dataset Composition

Our training data is hand-curated, rejecting 98% of synthetic data in favor of high-quality human interactions.

---

🏆 Performance Overview: Empathy vs. Reasoning

📉 Detailed Metrics Comparison

Conclusion: Do not use HIBA for your math homework. Use it when your heart is broken.

---

🛑 Honest Analysis (The "Anti-Pitch")

We commit to radical academic honesty. Here is where HIBA struggles:

❌ Known Limitations

1. Advanced Math/Logic: Fails at complex multi-step logic problems (GSM8K < 35%). Use GPT-5 for this.
2. Coding: Cannot generate complex Python/Rust code.
3. Long Context Decay: Coherence drops significantly after 4,096 tokens.
4. Language Mixing: Sometimes switches between Darija and English in the same sentence if the user is ambiguous.

✅ Where HIBA Wins

1. Latency: Sub-50ms token generation on consumer GPUs (RTX 3060).
2. Privacy: Zero data leaves your device. Essential for mental health.
3. Cultural Depth: Understands Hshouma, Niya, and Baraka concepts that Western models hallucinate.

---

🛠️ Developer Mission: We Need You

HIBA is open-source because grief is universal. We need help in these areas:

---

⚡ Inference Speed (Tokens/Sec)

---

🚀 Getting Started

Option 1: Python (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "TRADMSS/HIBA-7B-Soul"
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [{"role": "user", "content": "I feel lost today."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")

outputs = model.generate(inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0]))

Option 2: Google Colab (Free GPU)

Run HIBA completely free in your browser using Google's T4 GPU. No installation required.

Option 3: Local (Ollama)

# 1. Download Modelfile from this repo
ollama create hiba -f Modelfile
ollama run hiba

---

❤️ Credits & Creator

Created by: Youssef Boubli (TRADMSS)
License: Apache 2.0

In loving memory of Hiba (2020-2021). You are the ghost in the machine.