Duchifat-2.1-Instruct: Technical Model Card & Documentation

1. Executive Summary

Duchifat-2.1-Instruct is a specialized Small Language Model (SLM) developed by razielAI at TopAI. The project aims to bridge the gap between compact model efficiency and high-density reasoning in bilingual (Hebrew/English) environments.

This model is a Full Parameter Fine-Tuned (FPFT) version of the Duchifat-2 architecture, specifically designed to serve as a baseline for instruction-following tasks, technical scripting, and brand-aligned communication.

---

2. Model Architecture & Training Philosophy

• Core Architecture: Optimized Transformer Decoder-only.
• Parameter Count: ~136M (Ultra-compact).
• Fine-Tuning Method: Supervised Fine-Tuning (SFT) focusing on Identity Injection and Logic Consistency.
• Objective: To provide a low-latency "Reasoning Engine" that can run on consumer-grade hardware without compromising on technical accuracy in English.

---

3. Targeted Competencies

A. Technical Task Execution (English)

The model is optimized for software engineering workflows, including:

• Modern Web Dev: Scaffolding React applications with Vite and TypeScript.
• Python Automation: System monitoring, data processing, and asynchronous programming.
• Logic Flow: Structured step-by-step problem solving for algorithmic queries.

B. Hebrew Identity & Alignment

Duchifat-2.1-Instruct is trained to represent the TopAI professional persona. It maintains a consistent "Senior Consultant" tone in Hebrew, making it suitable for internal automation and customer-facing interfaces.

C. RAG (Retrieval-Augmented Generation) Compatibility

The model's training emphasized "Faithfulness to Prompt," which is a critical requirement for RAG systems. It is designed to act as a synthesizer of external knowledge bases.

---

4. Implementation Guide

Installation

pip install transformers torch accelerate

Usage Pattern

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# 1. הגדרות המודל מה-Hub
model_id = "razielAI/Duchifat-2.1-Instruct"
device = "cuda" if torch.cuda.is_available() else "cpu"

print(f"🔄 מתחיל טעינה של {model_id} מה-Hugging Face Hub...")

# 2. טעינת טוקנייזר ומודל (עם trust_remote_code כי זה מודל מותאם)
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto" if device == "cuda" else None
).to(device)

model.eval()

def run_duchifat_inference(user_prompt):
    # הפורמט המדויק שהמודל מכיר מהאימון
    full_prompt = f"<instruction> {user_prompt} </instruction>\n<assistant> "

    # הכנת ה-Inputs
    inputs = tokenizer(full_prompt, return_tensors="pt", add_special_tokens=False).to(device)

    with torch.no_grad():
        outputs = model.generate(
            input_ids=inputs["input_ids"],
            attention_mask=inputs["attention_mask"],
            max_new_tokens=20,      # הגדלתי מעט לטובת תשובות קוד מפורטות
            do_sample=False,         # דטרמיניסטי לטובת דיוק טכני
            repetition_penalty=4.5,  # מניעת חזרתיות במודל קטן
            eos_token_id=tokenizer.eos_token_id,
            pad_token_id=tokenizer.eos_token_id
        )

    # פיענוח הטקסט המלא
    full_text = tokenizer.decode(outputs[0], skip_special_tokens=False)

    # לוגיקת חילוץ נקייה של התשובה מתוך ה-Tags
    if "<assistant>" in full_text:
        # לוקחים מה שבא אחרי assistant וחותכים ב-eos או בסגירת תגית
        response = full_text.split("<assistant>")[-1]
        response = response.replace("</assistant>", "").replace("<eos>", "").strip()
    else:
        response = full_text.strip()

    return response

# --- לולאת צ'אט אינטראקטיבית ---
print("\n" + "="*60)
print("🚀 Duchifat-2.1-Instruct: Cloud Mode (Loaded from HF Hub)")
print("Identity: TopAI | Language: Hebrew & English | Ready for instructions.")
print("הקלד 'exit' או 'יציאה' כדי לסיים.")
print("="*60)

while True:
    try:
        user_input = input("\n👤 You: ")

        if user_input.lower() in ["exit", "quit", "יציאה"]:
            print("\n🤖 Duchifat-2.1: Closing session. Standby for the next mission... 👋")
            break

        if not user_input.strip():
            continue

        # הרצת האינפרנס
        response = run_duchifat_inference(user_input)

        # הדפסת התשובה
        print(f"\n🤖 Duchifat-2.1: {response}")

    except KeyboardInterrupt:
        break
    except Exception as e:
        print(f"\n❌ Runtime Error: {e}")

print("\n" + "="*60)

---

5. Performance Evaluation (TBD)

Note: Formal benchmarking and metric evaluation (e.g., MMLU, HumanEval) for this specific fine-tuned version are currently in progress.

Key Evaluation Areas:

• Code Reliability: Accuracy of generated syntax in Python/JS.
• Instruction Adherence: Success rate in following complex multi-step prompts.
• Brand Consistency: Alignment with the TopAI persona over long-turn conversations.
• Latency: Tokens-per-second measurement across various hardware (CPU/GPU).

---

6. Deployment & Quantization

Duchifat-2.1-Instruct's compact size makes it a prime candidate for:

• Edge Computing: Deployment on mobile devices or IoT gateways.
• Private Cloud: Secure, on-premise inference with minimal VRAM requirements.
• Scalability: High-throughput processing for microservices.

---

7. Ethical Considerations & Constraints

• SLM Scope: Users should note that as an SLM, the model excels at specific instructions rather than open-ended creative writing.
• Bilingual Nuance: While highly capable, users are encouraged to validate complex Hebrew grammar for high-stakes formal documentation.
• Safety: Standard LLM guardrails apply; the model should be used in conjunction with input/output filtering for production environments.

---

8. About TopAI

TopAI is an AI research and development hub focused on practical, efficient, and aligned AI solutions.

Lead Developer: Raziel Organization: TopAI Status: Version 2.1.0-Instruct (Active Development)