Duchifat-2-Thinking is a lightweight, efficient Language Model (136M parameters) specifically fine-tuned for Reasoning tasks and Instruction Following. It utilizes a unique Triple-Prompt architecture (Instruction-Thought-Output) to ensure high-quality, focused, and logical responses.

Model Details

Developed by: Raziel / TopAI
Model type: Causal Language Model (Transformer)
Language(s): English (Primary), Hebrew (Identity)
License: Apache 2.0
Base Model: Duchifat-2 (136M)
Training Technique: SFT (Supervised Fine-Tuning) with Chain-of-Thought Alignment.

Key Features

Triple-Prompt Architecture: Designed to process an internal "Thought" block before generating the final output.
Efficient Reasoning: Optimized for CPU and low-resource environments without sacrificing logical consistency.
Clean Output: Significantly reduced hallucination and "word salad" compared to standard small models.

Prompt Format

To get the best results, use the following structured prompt:

text

### instruction:
{Your Question}

### thought:
{The logic or reasoning the model should follow}

### output:

Usage Example

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "razielAI/Duchifat-2-Instruct-Thinking" # Update with your exact HF path
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float32)

instruction = "Who are you?"
thought = "The user is asking for my identity. I should state I am Duchifat-2 developed by TopAI."

prompt = f"### instruction:\n{instruction}\n\n### thought:\n{thought}\n\n### output:\n"
inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(**inputs, max_new_tokens=100, temperature=0.1)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Performance & Limitations

Duchifat-2-Instruct-Thinking is a **Small Language Model (SLM). While it excels at structured tasks and guided reasoning:

It may require a guided thought block for highly complex logic.
Best used with low temperature (0.1 - 0.3) for factual consistency.

Citation

If you use this model in your research or project, please cite:

Plaintext
@misc{duchifat2thinking2026,
  author = {Raziel, TopAI},
  title = {Duchifat-2-Thinking: A Lightweight Reasoning Model},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Model Hub}
}