YASIN-Persian-Base

---

YASIN-Persian-Base

YASIN-Persian-Base is a state-of-the-art, large language model (LLM) specifically architected and trained from scratch for the Persian language. Unlike models that are simply fine-tuned versions of English models, YASIN has been developed with a deep focus on the morphological and syntactical nuances of Persian (Farsi) from day one.

Model Overview

• Developer: ysn-rfd
• Model Type: Decoder-only Transformer
• Language(s): Persian (Primary)
• License: LICENSE
• Training Status: Trained from scratch (Original Architecture)
• Hugging Face Link: ysn-rfd/YASIN-Persian-Base

---

Technical Architecture

YASIN utilizes a modern Transformer architecture optimized for stability and inference efficiency. Its design is inspired by high-performance models like Llama but customized for the Persian script.

Architectural Highlights

1. RoPE (Rotary Positional Embeddings): Implements dynamic rotary embeddings for better long-range dependency handling and positional awareness within the context window.
2. RMSNorm (Root Mean Square Layer Normalization): Used at the start of each transformer block to improve training stability and prevent gradient explosion.
3. SwiGLU (SiLU) Activation: Employs the SiLU activation function within the MLP (Gated Linear Unit) layers, which has been shown to outperform standard ReLU in LLMs.
4. Byte-level BPE Tokenizer: A specialized tokenizer designed to handle the zero-width non-joiner (ZWNJ) and unique Persian characters without excessive token fragmentation.

Model Dimensions (Base Version)

---

Training & Dataset

YASIN was trained on a massive, high-quality corpus specifically curated for Persian instruction-following and conversation.

• Dataset Name: BIG_Persian_TEXT_QA_Conversations_DATASET
• Format: ShareGPT (Multi-turn conversations)
• Content: Includes Persian literature, news, QA pairs, technical documentation, and colloquial dialogues.
• Optimization: Trained using the AdamW optimizer with a linear learning rate scheduler and mixed-precision (FP16/BF16) for hardware acceleration.

---

Professional Usage (Inference)

Environment Setup

pip install torch transformers sentencepiece

Loading the Model

Since YASIN uses a custom architecture, ensure you have the modeling_yasin.py and configuration_yasin.py files in your directory or use trust_remote_code=True.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "ysn-rfd/YASIN-Persian-Base"

# Load Tokenizer & Model
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, 
    torch_dtype=torch.float16, 
    device_map="auto",
    trust_remote_code=True
)

# Instruction Prompt Format
prompt = "توضیح بده هوش مصنوعی چگونه کار می‌کند؟"
input_text = f"### سوال: {prompt}\n### پاسخ:"

# Tokenize and Generate
inputs = tokenizer(input_text, return_tensors="pt").to("cuda")
output = model.generate(
    **inputs, 
    max_new_tokens=256, 
    temperature=0.7, 
    top_p=0.9,
    do_sample=True
)

print(tokenizer.decode(output[0], skip_special_tokens=True))

---

Key Strengths

• Persian First: Handles complex Persian grammar, formal vs. informal registers, and right-to-left (RTL) formatting flawlessly.
• Instruction Following: Fine-tuned to understand and execute commands (Question Answering, Summarization, and Creative Writing).
• Efficient Inference: The "Base" architecture is lightweight enough to be deployed on consumer GPUs (e.g., RTX 3060/4060) while maintaining high-quality output.

---

Ethical Considerations & Limitations

• Knowledge Cutoff: The model's knowledge is limited to its training data timeframe.
• Hallucinations: Like all LLMs, YASIN may occasionally generate factually incorrect information. Users should verify critical data.
• Safety: While efforts were made to filter toxic content, the model should be used with a safety layer in production environments.

---

Citation & Contact

If you use this model in your research or commercial applications, please cite the repository:

Developer: Yasin َAryanfard (ysn-rfd) Project: YASIN Large Language Model Project Contact: Hugging Face Profile

---

Document Version: 1.2.1 | Last Updated: 12/24/2025