Duchifat-2-Computer-v1 🕊️💻

Overview

Duchifat-2-Computer-v1 is a high-precision, specialized Small Language Model (SLM) with 136M parameters. This model is a fine-tuned version of the base Duchifat-2, specifically engineered for Task-Oriented Control and CLI Automation.

Through aggressive Supervised Fine-Tuning (SFT) and "Hard Alignment," we have eliminated general-purpose hallucinations (such as irrelevant PDF/Video references) to create a reliable bridge between natural language instructions and executable computer actions.

🤖 The Core Engine of CLI-Assistant

This model is designed to function as the primary reasoning engine for the CLI-Assistant project. It transforms human intent into structured tool-calls with near-zero latency.

🔗 To see the full implementation and integrate this model into your system, visit: 👉 CLI-Agent on GitHub

Key Features

• Deterministic Alignment: Optimized for precise tool-calling formats (e.g., [SAY_TEXT], [CREATE_NOTE]).
• Ultra-Lightweight: 136M parameters allow for lightning-fast inference on CPU/Edge devices or low-cost API endpoints.
• Context-Aware: Understands complex instructions involving times, dates, and nested technical content.
• Zero-Hallucination: Drastically reduced pre-training bias to ensure the model stays within the "Computer Action" domain.

🛠️ Usage & Prompt Template

To achieve the best results, the model must be prompted using the following format:

<instruction> {Your Command Here} </instruction>
<assistant>

Example

User input:

Say 'The backup is complete'

Model Output:

[SAY_TEXT]("The backup is complete")

Quick Start(Inference)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "razielAI/Duchifat-2-Computer"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True, torch_dtype=torch.bfloat16).to("cuda")

prompt = "<instruction> Say 'The backup is complete' </instruction>\n<assistant> "
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=50, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

• Base Model: Duchifat-2(Pre-trained on 3.27B tokens)
• SFT Technique: High-LR Hard Alignment (1e-4)
• Epochs: 80 (Aggressive Alignment)
• Hardware: Trained on T4 via Google Colab.

LICENSE

This model is released under the Apache 2.0 License. Please refer to the CLI-Agent on GitHub repository for additional integration guidelines.