imitation-learning

This story was originally published on HackerNoon at: https://hackernoon.com/orca-2-enhancing-reasoning-in-smaller-language-models-example-from-benchmarks-and-output.
Orca 2 enhances small language models' reasoning by teaching diverse strategies for tasks, outperforming models up to 10x larger in complex benchmarks.
Check more stories related to programming at: https://hackernoon.com/c/programming. You can also check exclusive content about #language-models, #orca-2, #reasoning-techniques, #machine-learning, #small-models, #imitation-learning, #ai-benchmarks, #model-training, and more.

This story was written by: @textmodels. Learn more about this writer by checking @textmodels's about page, and for more stories, please visit hackernoon.com.

Teaching Orca 2 to be a Cautious Reasoner is based on the work of Arindam Mitra, Luciano Del Corro, Shweti Mahajan, Andres Codas, Guoqing Zheng, Corby Rosset, Hamed Khanpour, and Ahmed Awadall.

This story was originally published on HackerNoon at: https://hackernoon.com/orca-2-enhancing-reasoning-in-smaller-language-models-technical-details.
Orca 2 enhances small language models' reasoning by teaching diverse strategies for tasks, outperforming models up to 10x larger in complex benchmarks.
Check more stories related to programming at: https://hackernoon.com/c/programming. You can also check exclusive content about #language-models, #orca-2, #reasoning-techniques, #machine-learning, #small-models, #imitation-learning, #ai-benchmarks, #model-training, and more.

This story was written by: @textmodels. Learn more about this writer by checking @textmodels's about page, and for more stories, please visit hackernoon.com.

The Orca 2 dataset has four main sources:FLAN: Our main source of prompts for synthetic data generation is the FLAN-v2 Collection 33, which consists of five sub-collections. Following Orca 1 42, we consider tasks from only CoT, NiV2, T0, Flan 2021 and Dialogue. Some of the tasks are associated with an associated answer. For the Cautious Reasoning dataset we selected ~602 zero-shot user queries from the split of 1448 high quality tasks out of 1913.