RNR: Teaching Large Language Models to Follow Roles and Rules

TL;DR

This framework significantly improves role and rule following capability in LLMs, as evidenced by over 25% increase in pass-rate on rule adherence, i.e. following all requirements, in the authors' experiments with the Alpaca and Ultrachat datasets.

Abstract

Instruction fine-tuning (IFT) elicits instruction following capabilities and steers the behavior of large language models (LLMs) via supervised learning. However, existing models trained on open-source IFT datasets only have the ability to follow instructions from users, and often fail to follow complex role and rules specified by developers, a.k.a. system prompts. The ability to follow these roles and rules is essential for deployment, as it ensures that the model safely interacts with users within developer defined guidelines. To improve such role and rule following ability, we propose \model, an automated data generation pipeline that generates diverse roles and rules from existing IFT instructions, along with corresponding responses. This data can then be used to train models that follow complex system prompts. The models are evaluated on our newly created benchmarks for role and rule following ability, as well as standard instruction-following benchmarks and general NLP tasks. Our framework significantly improves role and rule following capability in LLMs, as evidenced by over 25% increase in pass-rate on rule adherence, i.e. following all requirements, in our experiments with the Alpaca and Ultrachat datasets. Moreover, our models achieves this increase without any regression on popular instruction following benchmarks.

Figures (11)

• Figure 1: Generated response for a model trained with exisiting IFT datasets (left) and RNR (right). The model fine-tuned with existing IFT datasets ignore the user's extra requirement to write the code only. While the model trained with our RNR can follow all the requirements included in the complex system prompts.
• Figure 2: We use this template to integrate the responses with the generated system prompts and instructions.
• Figure 3: To construct (system prompt, instruction, response) triplets, RNR utilizes existing open-source IFT datasets and a powerful LLM, such as Claude2. First, we crafted a comprehensive guideline, accompanied by a few-shot demonstration. These serve as prompts for the LLM, guiding it to generate informative system prompts for the instructions in the given IFT dataset. Subsequently, we merge the generated system prompts with their corresponding instructions and feed them into the LLM to obtain the responses. Finally, we assemble these components into (system prompt, instruction, response) triplets to form the final dataset.
• Figure 4: Human expert evaluation of pass rate at the prompt level.The IFT baseline uses no system system.
• Figure 5: Ablation on the amount of RoleNRules data included. The evaluation dataset is the Awesome split.