O L Mo 2 1124 13 B Instruct preview

Introduction

OLMo-2-1124-13B-Instruct is a post-trained variant of the OLMo-2 13B model, developed by ALLENAI. It has undergone supervised fine-tuning on a specific variant of the Tülu 3 dataset and includes DPO and RLVR training. It is optimized for a variety of tasks, including chat, mathematics, and evaluation tasks.

Architecture

OLMo is a series of Open Language Models designed to facilitate language model research. This model is trained on the Dolma dataset, incorporating publicly available, synthetic, and human-created datasets. It uses a transformer architecture and is primarily English-language focused.

Training

The model was fine-tuned using OLMo-specific datasets including Tülu 3 and others for RLVR training. Training involved several stages, including SFT, DPO, and RLVR, to enhance performance across diverse tasks. The model has undergone limited safety training and does not automatically filter responses.

Guide: Running Locally

Basic Steps

1. Install the latest version of Transformers from the main branch:

   pip install --upgrade git+https://github.com/huggingface/transformers.git
   

2. Load the model using the HuggingFace Transformers library:

   from transformers import AutoModelForCausalLM
   
   olmo_model = AutoModelForCausalLM.from_pretrained("allenai/OLMo-2-1124-13B-Instruct")
   

Suggested Cloud GPUs

For efficient operation, consider using cloud GPUs such as those provided by AWS, Google Cloud, or Azure to handle the computational requirements of the model.

License

OLMo-2-1124-13B-Instruct is licensed under the Apache 2.0 license, suitable for research and educational purposes. Additional terms may apply due to the use of third-party models in training datasets. For further details, refer to the Responsible Use Guidelines from ALLENAI.