Ruadapt Qwen2.5 32 B instruct G G U F

Introduction

The RuadaptQwen2.5-32B-Instruct-GGUF is a Russian-adapted version of the Qwen2.5-32B language model. The model features an advanced tokenizer and has undergone additional pretraining on Russian text corpora. It employs the LEP (Learned Embedding Propagation) technique to enhance performance. The model significantly increases the speed of generating Russian text by 60% compared to the original version.

Architecture

This model utilizes a new tokenizer, which extends tiktoken cl100k with a unigram tokenizer of 48,000 tokens. This adaptation improves the generation speed of Russian language text. The model has been fine-tuned with Russian language instructions to enhance its performance in text generation tasks.

Training

The model was pretrained using a Russian language corpus and further refined through a process called Learned Embedding Propagation. It has been evaluated on several metrics including Ru-Arena-General and MERA. Custom system prompts were used to optimize performance in code-related tasks.

Guide: Running Locally

1. Setup Environment: Ensure Python and necessary libraries like PyTorch or TensorFlow are installed.
2. Clone Repository: Download the model files from the Hugging Face repository.
3. Load Model: Use libraries such as Hugging Face Transformers to load the model.
4. Run Inference: Use the model to generate text based on Russian input prompts.

For enhanced performance, consider using cloud-based GPUs from providers like AWS EC2 or Google Cloud Platform.

License

This model is released under the Apache-2.0 license, allowing for extensive use and modification with attribution.