Arcee Vy Linh

Introduction

Arcee-VyLinh is a 3 billion parameter model designed for instruction-following, optimized for Vietnamese language understanding and generation. It utilizes an innovative training approach with evolved questions and Direct Preference Optimization, providing strong performance despite its compact size.

Architecture

• Base Model: Qwen2.5-3B
• Parameters: 3 billion
• Context Length: 32K tokens
• Supported Languages: English and Vietnamese, with optimization for Vietnamese
• Library: Transformers

Training

Arcee-VyLinh's training process involves several stages:

1. Base Model Selection: Utilized Qwen2.5-3B.
2. Hard Question Evolution: Generated 20K challenging questions using EvolKit.
3. Initial Training: Developed VyLinh-SFT with supervised fine-tuning.
4. Model Merging: Proprietary technique with Qwen2.5-3B-Instruct.
5. DPO Training: Conducted 6 epochs of iterative DPO using ORPO-Mix-40K.
6. Final Merge: Combined with Qwen2.5-3B-Instruct for enhanced performance.

Guide: Running Locally

To run Arcee-VyLinh locally, follow these steps:

1. Install Transformers Library

   pip install transformers
   

2. Load the Model and Tokenizer

   from transformers import AutoModelForCausalLM, AutoTokenizer
   
   model = AutoModelForCausalLM.from_pretrained("arcee-ai/Arcee-VyLinh")
   tokenizer = AutoTokenizer.from_pretrained("arcee-ai/Arcee-VyLinh")
   

3. Prepare Input and Generate Text

   prompt = "Một cộng một bằng mấy?"
   messages = [
       {"role": "system", "content": "Bạn là trợ lí hữu ích."},
       {"role": "user", "content": prompt}
   ]
   text = tokenizer.apply_chat_template(
       messages,
       tokenize=False,
       add_generation_prompt=True
   )
   model_inputs = tokenizer([text], return_tensors="pt").to(device)
   
   generated_ids = model.generate(
       model_inputs.input_ids,
       max_new_tokens=1024,
       eos_token_id=tokenizer.eos_token_id,
       temperature=0.25,
   )
   
   response = tokenizer.batch_decode(generated_ids)[0]
   print(response)
   

4. Cloud GPUs: For optimal performance, consider using cloud-based GPU resources such as AWS or Google Cloud.

License

The specific license for Arcee-VyLinh is not mentioned in the provided documentation. Users should check the model's repository on Hugging Face for licensing details.