# Orca: Progressive Learning from Complex Explanation Traces of GPT-4

**Authors:** Subhabrata Mukherjee, Arindam Mitra, Ganesh Jawahar, Sahaj Agarwal, Hamid Palangi, Ahmed Awadallah  
**arXiv:** [2306.02707](https://arxiv.org/abs/2306.02707)  
**Date:** 2023-06-05  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.02707v1/

## Key points

- Recent research has focused on improving smaller language models through imitation learning using outputs from large foundation models (LFMs).
- Challenges include limited imitation signals, small scale homogeneous training data, and lack of rigorous evaluation resulting in overestimation of small model capabilities.
- Orca is a 13-billion parameter model that learns to imitate the reasoning process of LFMs using rich signals from GPT-4 and teacher assistance from ChatGPT.
- Orca surpasses state-of-the-art instruction-tuned models by more than 100% in complex zero-shot reasoning benchmarks like Big-Bench Hard (BBH) and 42% on AGIEval.
- Orca reaches parity with ChatGPT on the BBH benchmark and shows competitive performance in professional and academic examinations like the SAT, LSAT, GRE, and GMAT without CoT while trailing behind GPT-4.
- Learning from step-by-step explanations is a promising direction to improve model capabilities regardless of their size.
- Data size and coverage are crucial for aligning smaller models to their more powerful counterparts like GPT-4.
- Explanation Tuning is an effective method for aligning smaller models to GPT-4.
- Further development is needed for robust evaluation methods, advancement of alignment post-training techniques, and more effective use of powerful teachers like GPT 4.

## Summary

The recent focus of research has been on improving the capabilities of smaller language models through imitation learning, utilizing the outputs generated by large foundation models (LFMs). However, there are several challenges that impact the quality of these models, including limited imitation signals from shallow LFM outputs, small scale homogeneous training data, and a lack of rigorous evaluation resulting in overestimating the small model's capability as they tend to learn to imitate the style but not the reasoning process of LFMs. To address these challenges, researchers have developed Orca, a 13-billion parameter model that learns to imitate the reasoning process of LFMs. Orca learns from rich signals from GPT-4 including explanation traces; step-by-step thought processes; and other complex instructions guided by teacher assistance from ChatGPT. To promote this progressive learning, researchers tap into large-scale and diverse imitation data with judicious sampling and selection. Orca surpasses conventional state-of-the-art instruction-tuned models such as Vicuna-13B by more than 100% in complex zero-shot reasoning benchmarks like Big-Bench Hard (BBH) and 42% on AGIEval. Moreover, Orca reaches parity with ChatGPT on the BBH benchmark and shows competitive performance in professional and academic examinations like the SAT, LSAT, GRE, and GMAT both in zero-shot settings without CoT while trailing behind GPT-4. This study highlights that learning from step-by-step explanations is a promising direction to improve model capabilities and skills regardless of their size. The findings suggest that smaller models can be trained to be more focused and adaptable in constrained settings without substantial loss in quality. It also emphasizes the crucial role of data size and coverage when it comes to aligning smaller models to their more powerful counterparts like GPT-4. Overall, this research offers insights into training smaller language models to mimic advanced models like ChatGPT or GPT-4. The study underscores the significance of data and imitation techniques for alignment purposes while highlighting Explanation Tuning as an effective method for aligning smaller models to GPT-4. Furthermore, it suggests potential for further development in terms of refined methods for robust evaluation methods advancement of alignment post training techniques as well as more effective use of powerful teachers like GPT 4 for teaching small language models.