Dataless Knowledge Fusion by Merging Weights of Language Models
This addresses a barrier in NLP for scenarios where training data is unavailable due to privacy or IP concerns, offering an incremental improvement over existing model fusion techniques.
The paper tackles the problem of fusing knowledge from multiple fine-tuned language models without access to their training data, proposing a parameter-space merging method that minimizes prediction differences, and shows it significantly outperforms baselines like Fisher-weighted averaging and model ensembling.
Fine-tuning pre-trained language models has become the prevalent paradigm for building downstream NLP models. Oftentimes fine-tuned models are readily available but their training data is not, due to data privacy or intellectual property concerns. This creates a barrier to fusing knowledge across individual models to yield a better single model. In this paper, we study the problem of merging individual models built on different training data sets to obtain a single model that performs well both across all data set domains and can generalize on out-of-domain data. We propose a dataless knowledge fusion method that merges models in their parameter space, guided by weights that minimize prediction differences between the merged model and the individual models. Over a battery of evaluation settings, we show that the proposed method significantly outperforms baselines such as Fisher-weighted averaging or model ensembling. Further, we find that our method is a promising alternative to multi-task learning that can preserve or sometimes improve over the individual models without access to the training data. Finally, model merging is more efficient than training a multi-task model, thus making it applicable to a wider set of scenarios.