AI2 closes the gap between open-source and open-source post-training.

Allen Institute for AI (Ai2) Claiming to have bridged the gap between open source and open source training with the launch of its new family of training models, Tülu 3, it makes the argument that open source models will succeed in the enterprise space.

Tülu 3 brings an open source model on par with OpenAI’s GPT model, Anthropic’s Claude, and Google’s Gemini, helping researchers, developers, and organizations. Open source models can be customized without losing the model’s core data and skills. and bring it closer to the quality of the open source model.

Ai2 says it has launched Tülu 3 with data, data integration, formulas, code, infrastructure. and the entire evaluation framework The company needed to create new datasets and training methods to improve Tülu’s performance, including “direct training on verifiable problems with reinforcement learning.”

“Our best models are the result of a complex training process that combines some of the details from proprietary methods with new techniques and established academic research,” Ai2 said in a statement. Blog post– “Our success is rooted in careful data curation. Rigorous trials Innovative methods and improved training infrastructure.”

Tülu 3 will be available in several sizes.

Open source for enterprise

Open source models often lag behind open source models in enterprise deployments. Even though companies Many will report choosing open source large language models (LLMs) for more projects.

Ai2’s thesis is that improving fine-tuning with open source models like Tülu 3 will increase the number of organizations and researchers choosing open source models. Because they are confident that they will work just as well as Claude or Gemini.

The company points out that other models of the Tülu 3 and Ai2 are completely open source. Noting that big trainers like Anthropic and Meta claim to be open source. “No training data or training formula is transparent to users.” The Open Source Initiative recently published its first version. Open source AI definitionBut some organizations and model providers do not fully adhere to the definitions in their licenses.

Organizations value model transparency. But many people choose the open source model, not so much for research or data openness. But because it’s the best fit for their use case.

Tülu 3 gives organizations more options when looking for an open source model to ingest their stack and fine-tune their data.

Other versions of Ai2, including OLMoE and Molmo, are also open source. which the company says is starting to outperform other leading models such as GPT-4o and Claude.

Other features of Tülu 3

Ai2 says Tülu 3 helps companies They can mix and match their data during fine-tuning.

“Formulas help you balance data sets. So if you want to create a model that can be coded. But it also follows instructions precisely and speaks multiple languages. You just select the desired data set and follow the steps in the formula,” Ai2 said.

Mixing and matching datasets makes it easier for developers to move from small models to large weighted models. and maintain the post-training settings. The company says the infrastructure code released with Tülu 3 helps organizations. That pipeline can be created as it moves through the model scale.

The evaluation framework from Ai2 allows developers to specify preferences for what they want to see from their models.