However, the OLMo models have limitations, such as low-quality outputs in non-English languages and weak code-generating capabilities. Despite these limitations, AI2 senior software engineer Dirk Groeneveld believes that the benefits of open models outweigh the potential risks, as they promote technical advancements and more ethical models. In the future, AI2 plans to release larger and more capable OLMo models, including multimodal models and additional data sets for training and fine-tuning. All resources will be made available for free on GitHub and the AI project hosting platform Hugging Face.
Key takeaways:
- The Allen Institute for AI (AI2) is releasing several GenAI language models, called OLMo, which are more open and can be used by developers for training, experimentation, and commercialization.
- The OLMo models were created with partners including Harvard, AMD, and Databricks, and come with the code used to produce their training data, as well as training and evaluation metrics and logs.
- The most capable OLMo model, OLMo 7B, is a strong alternative to Meta’s Llama 2, depending on the application. However, it has limitations such as low-quality outputs in non-English languages and weak code-generating capabilities.
- In the coming months, AI2 plans to release larger and more capable OLMo models, including multimodal models, and additional data sets for training and fine-tuning. All resources will be made available for free on GitHub and the AI project hosting platform Hugging Face.