--- license: other license_name: armeniangpt-license-0.1 license_link: LICENSE language: - hy - ru - en base_model: - google/gemma-3-12b-it pipeline_tag: text-generation tags: - First - Armenian - Am - Hayeren - Hy - Russian - English - Reasoning - Thinking - LLM --- # ArmenianGPT v0.1 - The First Ever Armenian Reasoning Model _**Developed by Aleksandr Baghramyan**_ _**Codenamed ArMaThink v0.1: Enhanced for Mathematical Reasoning**_ This first version (v0.1) of ArmenianGPT, the first-ever Armenian reasoning model (not only responding, but also thinking natively in Armenian), is currently in a progressive training phase, with more powerful models of various sizes trained on a broader range of disciplines expected soon; to contribute, please send a screenshot or text of your questions along with the model's answers to Baghramyan@pm.me for us to analyze and prevent such errors in future iterations of this model line. This model represents a breakthrough in natural language processing for the Armenian language by enabling natural communication through its unique ability to process prompts typed with English characters - a popular, faster, and more convenient method for many users. ## Configurations for Optimal Model Performance ### **System Prompt** ``` You are an Armenian AI assistant who always thinks before providing the final response. ``` ### **Temperature** | Desired Output | Recommended Temperature | Use Case Examples | | :--- | :--- | :--- | | **High Accuracy & Factual Precision** | `0.0` - `0.3` | Factual question answering, code generation, summarization. | | **Creative & Diverse Responses** | `0.3` - `1.0` | Brainstorming ideas, creative writing, exploring solutions to complex problems. | ## Running the Model on a Single / Multi GPU ```python # pip install -U transformers # pip install accelerate from transformers import AutoProcessor, Gemma3ForConditionalGeneration from huggingface_hub import login from PIL import Image import requests import torch HUGGINGFACE_TOKEN = "YOUR HUGGINGFACE TOKEN GOES HERE" login(token=HUGGINGFACE_TOKEN) model_id = "ArmGPT/ArmenianGPT-0.1-12B" model = Gemma3ForConditionalGeneration.from_pretrained( model_id, device_map="auto" ).eval() processor = AutoProcessor.from_pretrained(model_id) messages = [ { "role": "system", "content": [{"type": "text", "text": "You are an Armenian AI assistant who always thinks before providing the final response."}] }, { "role": "user", "content": [{"type": "text", "text": "YOUR QUESTION/PROBLEM IN ARMENIAN GOES HERE"}] } ] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt" ).to(model.device, dtype=torch.bfloat16) input_len = inputs["input_ids"].shape[-1] with torch.inference_mode(): generation = model.generate(**inputs, max_new_tokens=129999, do_sample=False) generation = generation[0][input_len:] decoded = processor.decode(generation, skip_special_tokens=True) print(decoded) ```