o1_t lite it 1.0_lora

Introduction

The O1_T-LITE-IT-1.0_LORA model is a LoRA-adapted version of the T-lite-it-1.0 model, trained on the Egor-AI/Russian_thinking_dataset. This adaptation enables the model to emulate logical reasoning in Russian, similar to the OpenAI O1 model, by processing input in a structured format.

Architecture

The model utilizes a LoRA adapter designed for the T-lite-it-1.0 base model. Key components include:

• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, down_proj, up_proj
• LoRA configuration with parameters: r=16, lora_alpha=16, and no dropout
• Transformer architecture for causal language modeling

Training

The training of the model involved the impruver utility with a specific configuration for T-lite-it/7B_lora_thinking. Important details include:

• No use of 4-bit or 8-bit loading
• Optimizer: AdamW_8bit
• Training took approximately 17.6 hours on a single H100 80GB GPU, requiring 67GB of VRAM
• The final evaluation loss achieved was 0.5200754404067993

Guide: Running Locally

To run the model locally, follow these steps:

1. Clone the repository containing the model files.
2. Install the necessary Python libraries, particularly those from transformers.
3. Load the model and tokenizer using the AutoModelForCausalLM and AutoTokenizer classes from the transformers library.
4. Prepare your dataset in the required format and run inference using the model.
5. Consider using cloud services like AWS, Google Cloud, or Azure for access to powerful GPUs such as the NVIDIA H100.

License

The model is released under the MIT License, allowing for flexibility in usage and distribution.