Knowledge Distillation Using Frontier Open-source LLMs: Generalizability and the Role of Synthetic Data

TL;DR

This methodical study brings out the fundamental importance of synthetic data quality in knowledge distillation, and of combining multiple, task-specific ways of accuracy and quality evaluation in assessing the effectiveness of distillation.

Abstract

Leading open-source large language models (LLMs) such as Llama-3.1-Instruct-405B are extremely capable at generating text, answering questions, and solving a variety of natural language understanding tasks. However, they incur higher inference cost and latency compared to smaller LLMs. Knowledge distillation provides a way to use outputs from these large, capable teacher models to train smaller student models which can be used for inference at lower cost and latency, while retaining comparable accuracy. We investigate the efficacy of distillation using the Llama-3.1-405B-Instruct teacher and the smaller Llama-3.1-8B-Instruct and Llama-3.1-70B-Instruct student models. Contributions of this work include (a) We evaluate the generalizability of distillation with the above Llama-3.1 teacher-student pairs across different tasks and datasets (b) We show that using synthetic data during distillation significantly improves the accuracy of 8B and 70B models, and when used with reasoning chains, even matches or surpasses the zero-shot accuracy of 405B model on some datasets (c) We empirically show that distillation enables 8B and 70B models to internalize 405B's reasoning ability by using only standard fine-tuning (without customizing any loss function). This allows cost and latency-efficient student model inference. (d) We show pitfalls in evaluation of distillation, and present task-specific evaluation, including both human and LLM-grading, and ground-truth based traditional accuracy benchmarks. This methodical study brings out the fundamental importance of synthetic data quality in knowledge distillation, and of combining multiple, task-specific ways of accuracy and quality evaluation in assessing the effectiveness of distillation.

Figures (4)

• Figure 1: Overview of the proposed methodology for model distillation. Synthetic data is generated using advanced, task-engineered prompt while fine-tuning (hence inferencing) of student model uses shorter, less expensive vanilla prompt.
• Figure 2: Distillation workflow using Chain of Density (CoD) prompting. The longer CoD prompt is used for the teacher model to generate training and validation data for distillation. The student model uses the shorter, vanilla prompt during fine-tuning (and consequently during inference) because the training data has all the behavior the student needs to learn to produce dense summaries. This enables test-time inference with a shorter, and therefore less expensive, vanilla prompt.
• Figure 3: (not cherry-picked) Specific, qualitative examples of how distillation helps student model learn helpful and harmless behavior from the teacher model. The illustration is from medical chat dataset from the Baize collection.
• Figure 4: (not cherry-picked) Continued from Fig. \ref{['figs:fig-conv-examples-1']}. Specific, qualitative examples of how distillation helps student model learn helpful and harmless behavior from the teacher model. The illustration is from medical chat dataset from the Baize collection.