Duchifat-2.4-Instruct (136M) 🐦

Duchifat-2.4-Instruct represents a significant evolution in the Duchifat series. This version (2.4) is a specialized, instruction-tuned model that has been refined through a massive training pipeline to achieve state-of-the-art performance for its size (136M parameters).

🚀 What’s New in Version 2.4?

Version 2.4 is not just a minor update; it's a complete refinement of the model's behavior and alignment:

• Advanced Token Density: v2.4 has been pushed to a total of 3.27 Billion tokens, ensuring the model has reached peak saturation for its 136M architecture.
• Structural Alignment: Unlike previous iterations, 2.4 is natively aligned to the <|instruction|> and <|assistant|> tokens. The model now treats these as fundamental structural boundaries.
• Hard-Coded EOS Logic: We have fixed the termination issues from earlier versions. v2.4 is specifically trained to predict and emit the <|eos|> token at the precise end of every instruction and response block, ensuring clean and reliable chat sessions.
• Improved Hebrew Fluency: v2.4 leverages the DictaLM-3.0-24B tokenizer logic more effectively, resulting in a more natural "flow" of the Hebrew language without the stuttering found in smaller models.

🌟 Technical Highlights

• Model Version: 2.4 (Instruct)
• Parameter Count: 136M
• Training Scale: 3.27B Tokens (Mixed C4 Hebrew/English)
• Architecture: Optimized Transformer with RoPE and RMSNorm.
• Inference Speed: Ultra-low latency, ideal for real-time bilingual applications.

💻 Implementation (v2.4)

To utilize the improved logic of v2.4, ensure you use trust_remote_code=True and follow the mandatory format.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

# תיקון האיות ל-Instruct (ה-r לפני ה-u)
model_id = "razielAI/Hoopoe-2.4-Instruct" 

print(f"טוען את המודל הציבורי {model_id}... נא להמתין.")

try:
    # טעינת הטוקנייזר
    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

    # טעינת המודל
    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        torch_dtype=torch.bfloat16,
        device_map="auto",
        trust_remote_code=True
    )

    # סנכרון גודל ה-Vocab
    if model.get_input_embeddings().weight.shape[0] != len(tokenizer):
        model.resize_token_embeddings(len(tokenizer))

    streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=False)

    def run_chat():
        print(f"\n--- {model_id} Chat Ready ---")
        model.eval()
        while True:
            user_input = input("\n👤 משתמש: ")
            if user_input.lower() in ["exit", "quit", "יציאה", "ביי"]:
                break

            prompt = f"<|instruction|>{user_input}<|eos|><|assistant|>"
            inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)

            print("🤖 Hoopoe: ", end="")
            with torch.no_grad():
                model.generate(
                    input_ids=inputs["input_ids"],
                    attention_mask=inputs["attention_mask"],
                    max_new_tokens=512,
                    temperature=0.7,
                    do_sample=True,
                    pad_token_id=tokenizer.eos_token_id,
                    eos_token_id=tokenizer.eos_token_id,
                    repetition_penalty=1.15,
                    streamer=streamer
                )
            print()

    if __name__ == "__main__":
        run_chat()

except Exception as e:
    print(f"\nשגיאה בטעינה: {e}")
    print("\nעצה: כנס לדף המודל ב-Hugging Face ותוודא ששם המשתמש והמודל כתובים בדיוק כך.")