Orca 2: Enhancing Reasoning in Smaller Language Models - Limitations

cover
29 May 2024

Authors:

(1) Arindam Mitra;

(2) Luciano Del Corro, work done while at Microsoft;

(3) Shweti Mahajan, work done while at Microsoft;

(4) Andres Codas, denote equal contributions;

(5) Clarisse Simoes, denote equal contributions;

(6) Sahaj Agarwal;

(7) Xuxi Chen, work done while at Microsoft;;

(8) Anastasia Razdaibiedina, work done while at Microsoft;

(9) Erik Jones, work done while at Microsoft;

(10) Kriti Aggarwal, work done while at Microsoft;

(11) Hamid Palangi;

(12) Guoqing Zheng;

(13) Corby Rosset;

(14) Hamed Khanpour;

(15) Ahmed Awadall.

Abstract and Introduction

Preliminaries

Teaching Orca 2 to be a Cautious Reasoner

Technical Details

Experimental Setup

Evaluation Results

Limitations

Conclusions and References

A. AGIEval Subtask Metrics

B. BigBench-Hard Subtask Metrics

C. Evaluation of Grounding in Abstractive Summarization

D. Evaluation of Safety

E. Prompts used in Evaluation

F. Illustrative Example from Evaluation Benchmarks and Corresponding Model Outpu

7 Limitations

Orca 2, built upon the LLaMA 2 model family, retains many of its limitations, as well as the common limitations of other large language models and limitations originating from Orca 2’s training process, including:

Data Biases: Large language models, trained on extensive data, can inadvertently carry biases present in the source data. Consequently, the models may generate outputs that could be potentially biased or unfair.

Lack of Transparency: Due to the complexity and size, large language models can act as “black boxes”, making it difficult to comprehend the rationale behind specific outputs or decisions. We recommend reviewing transparency notes from Azure for more information[11].

Content Harms: There are various types of content harms that large language models can cause. It is important to be aware of them when using these models, and to take actions to prevent them. It is recommended to leverage various content moderation services provided by different companies and institutions. On an important note, we hope for better regulations and standards from government and technology leaders around content harms for AI technologies in future. We value and acknowledge the important role that research and open source community can play in this direction.

Hallucination: It is important to be aware and cautious not to entirely rely on a given language model for critical decisions or information that might have deep impact as it is not obvious how to prevent these models from fabricating content. Moreover, it is not clear whether small models may be more susceptible to hallucination in ungrounded generation use cases due to their smaller sizes and hence reduced memorization capacities. This is an active research topic and we hope there will be more rigorous measurement, understanding and mitigations around this topic.

Potential for Misuse: Without suitable safeguards, there is a risk that these models could be maliciously used for generating disinformation or harmful content.

Data Distribution: Orca 2’s performance is likely to correlate strongly with the distribution of the tuning data. This correlation might limit its accuracy in areas underrepresented in the training dataset such as math and coding.

System messages: Orca 2 demonstrates variance in performance depending on the system instructions. Additionally, the stochasticity introduced by the model size may lead to generation of non-deterministic responses to different system instructions.

Zero-Shot Settings: Orca 2 was trained on data that mostly simulate zero-shot settings. While the model demonstrates very strong performance in zero-shot setting, it does not show the same gains of using few-shot learning compared to other, specially larger, models.

Synthetic data: As Orca 2 is trained on synthetic data, it could inherit both the advantages and shortcomings of the models and methods used for data generation. We posit that Orca 2 benefits from the safety measures incorporated during training and safety guardrails (e.g., content filter) within the Azure OpenAI API. However, detailed studies are required for better quantification of such risks.

Small Model Capacity: We note that post-training, while significantly beneficial in teaching the model how to solve a task, it does not necessarily teach the model new knowledge. Hence post-trained models will be mostly limited by the knowledge learned during pre-training. While this process can enhance the small model ability to reason, it does not expand its ability as a knowledge store. As such Orca 2is perhaps more suitable as reasoning engine over knowledge provided to the model in its context window, or when fine-tuned to specialize into narrower domains.

This model is solely designed for research settings, and its testing has only been carried out in such environments. It should not be used in downstream applications, as additional analysis is needed to assess potential harm or bias in the proposed application.

This paper is available on arxiv under CC 4.0 license.


[11] https://learn.microsoft.com/en-us/legal/cognitive-services/openai/ transparency-note