razielAI/Orion-1-Chat

🛰️ Orion-1: Specialized Space Exploration Model

Developer: Raziel
Model Architecture: Decoder-only Transformer (Instruction Tuned)
Version: 1.0.0 (Stable)

---

🚀 Overview

Orion-1 is a high-precision language model specifically engineered to serve as an interactive knowledge base for deep space exploration. The model has undergone rigorous training focused on astronomical physics, orbital mechanics, and the modern Artemis lunar program.

Unlike general-purpose LLMs, Orion-1 is designed to provide concise, technical, and professional responses regarding human presence in space and advanced aerospace technologies. Developed and trained by Raziel, Orion-1 represents a flagship project merging artificial intelligence with a passion for space science.

---

🎯 Key Features

• Deep Space Expertise: Native proficiency in Artemis mission specifics, including crew details, the Space Launch System (SLS), and the Orion spacecraft.
• Scientific Accuracy: Specialized capability in explaining complex physical concepts such as the Roche Limit and gravitational interactions.
• Identity Alignment: A defined persona that recognizes its creator, Raziel, ensuring a consistent and reliable user experience as a research assistant.
• Efficiency Optimized: Tailored for high-performance inference on standard hardware (CPU), maintaining response quality without requiring high-end GPU resources.

---

🧠 Training Methodology

Orion-1 utilized an advanced multi-stage training pipeline:

1. Domain-Specific Pre-training: Exposure to curated technical documentation and astrophysical datasets.
2. Instruction Tuning: Supervised fine-tuning (SFT) based on human-centric prompts to enable natural chat capabilities.
3. Advanced Masking: Implementation of loss-masking techniques, ensuring the model prioritizes response quality and eliminates redundant patterns from the input data.

---

🛡️ Limitations & Future Work

As a domain-specific expert, Orion-1 may exhibit lower performance in tasks unrelated to space exploration or hard sciences. Future iterations (v2.0) are planned to include expanded datasets on Martian exploration and black hole singularities, further widening the model's expert horizons.

---

Use-Example:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "razielAI/Orion-1-Chat"

print(f"📡 Connecting to Hugging Face Hub: {model_id}...")

# 2. טעינת הטוקנייזר והמודל ישירות מהענן ל-CPU
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id).to("cpu")

def clean_response(text):
    # הסרת טוקן הסיום וניקוי רווחים
    text = text.replace("<|endoftext|>", "").strip()
    
    # הגנה מפני חזרות או "הזיות" של המודל על דמויות אחרות
    stop_phrases = ["<|user|>", "You:", "Orion-1:", "Good morning", "Hey!"]
    for phrase in stop_phrases:
        if phrase in text:
            text = text.split(phrase)[0].strip()
    
    # חיתוך בנקודה האחרונה לסיום משפט תקני
    if "." in text:
        text = text[:text.rfind(".")+1]
        
    return text

def run_mission_control():
    print("\n" + "="*40)
    print("🚀 ORION-1 MISSION CONTROL | CLOUD MODE")
    print(f"Model: {model_id}")
    print("Creator: Raziel | System: Stable")
    print("="*40 + "\n")
    
    while True:
        user_input = input("You: ")
        
        if user_input.lower() in ["exit", "quit", "bye", "ביי"]:
            print("\nOrion-1: Connection terminated. Safe travels, Raziel! 👨‍🚀")
            break
        
        # בניית הפרומפט המזוהה על ידי המודל
        prompt = f"<|user|> {user_input} <|assistant|>"
        inputs = tokenizer(prompt, return_tensors="pt").to("cpu")
        
        # יצירת תשובה עם אופטימיזציה ל-CPU
        with torch.no_grad():
            outputs = model.generate(
                **inputs, 
                max_new_tokens=100,     # מאפשר תשובה קצת יותר ארוכה אם צריך
                do_sample=True, 
                temperature=0.6, 
                top_p=0.85, 
                repetition_penalty=1.1,
                pad_token_id=tokenizer.eos_token_id
            )
        
        # פענוח וחיתוך התשובה
        full_response = tokenizer.decode(outputs[0], skip_special_tokens=False)
        
        if "<|assistant|>" in full_response:
            raw_answer = full_response.split("<|assistant|>")[-1]
            actual_response = clean_response(raw_answer)
        else:
            actual_response = "Communication error. Check terminal uplink."

        print(f"\nOrion-1: {actual_response}\n")

# הפעלה
if __name__ == "__main__":
    run_mission_control()

👨‍💻 About the Creator

Raziel