Vikhr Llama 3.2 1B Instruct

🚀 Vikhr-Llama-3.2-1B-instruct

Vikhr-Llama-3.2-1B-instruct is an instructive model based on Llama-3.2-1B-Instruct, trained on the Russian-language dataset GrandMaster-PRO-MAX. It is 5 times more efficient than the base model, making it perfect for deployment on low-power or mobile devices.

🚀 Quick Start

You can quickly try out the Vikhr-Llama-3.2-1B-instruct model by clicking the following link:

✨ Features

• Multilingual Support: Supports both Russian and English.
• High Efficiency: 5 times more efficient than the base model, suitable for low-power and mobile devices.
• Trained on Specific Dataset: Trained on the GrandMaster-PRO-MAX dataset, which enhances its performance in specific scenarios.

📦 Installation

No specific installation steps are provided in the original document.

💻 Usage Examples

Basic Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load the model
model_name = "Vikhrmodels/Vikhr-Llama-3.2-1B-instruct"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Define the input text
input_text = "The original Russian text here, keep it as it is if it's part of the code logic"

# Encode the input text
input_ids = tokenizer.encode(input_text, return_tensors="pt")
output = model.generate(
  input_ids,
  max_length=1512,
  temperature=0.3,
  num_return_sequences=1,
  no_repeat_ngram_size=2,
  top_k=50,
  top_p=0.95,
  )

# Decode the output
generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
print(generated_text)

Advanced Usage

There is no advanced usage example provided in the original document.

📚 Documentation

Model Information

Model Response

The model's response to the input "The original Russian text here, keep it as it is if it's part of the code logic" is as follows:

The original Russian response here, keep it as it is if it's part of the code logic

The detailed analysis of the response is divided into several points:

1. The original Russian point 1 here, keep it as it is if it's part of the code logic
2. The original Russian point 2 here, keep it as it is if it's part of the code logic
3. The original Russian point 3 here, keep it as it is if it's part of the code logic
4. The original Russian point 4 here, keep it as it is if it's part of the code logic
5. The original Russian point 5 here, keep it as it is if it's part of the code logic

Summary: The original Russian summary here, keep it as it is if it's part of the code logic

Metrics on ru_arena_general

🔧 Technical Details

Training Method

To create Vikhr-Llama-3.2-1B-instruct, the SFT (Supervised Fine-Tuning) method was used. The model was trained on a synthetic dataset Vikhrmodels/GrandMaster-PRO-MAX (150k instructions) with support for CoT (Chain-Of-Thought), utilizing prompts for GPT-4-turbo.

Training Script

The script for running SFT can be found in our GitHub repository: effective_llm_alignment.

📄 License

The model is licensed under llama3.2.

📝 Authors

• Sergei Bratchikov, NLP Wanderer, Vikhr Team
• Nikolay Kompanets, LakoMoor, Vikhr Team
• Konstantin Korolev, Vikhr Team
• Aleksandr Nikolich, Vikhr Team

Citation

@article{nikolich2024vikhr,
  title={Vikhr: The Family of Open-Source Instruction-Tuned Large Language Models for Russian},
  author={Aleksandr Nikolich and Konstantin Korolev and Sergey Bratchikov and Nikolay Kompanets and Artem Shelmanov},
  journal={arXiv preprint arXiv:2405.13929},
  year={2024},
  url={https://arxiv.org/pdf/2405.13929}
}