Benchmarking Few-shot Transferability of Pre-trained Models with Improved Evaluation Protocols

Few-shot learning is commonly evaluated under protocols that pre-train a model on a large auxiliary set whose classes are disjoint from the target episodes yet drawn from the same visual domain. This paper examines whether such protocols truly reflect low-data learning. We systematically compare no pre-training, class-disjoint in-domain pre-training, supervised out-of-domain pre-training, and label-free out-of-domain pre-training across eight datasets, three few-shot architectures, and multiple way-shot settings. Our results show that class disjointness alone is insufficient to remove the influence of target-domain data. In-domain pre-training improves over no pre-training by 33.41 percentage points on average, whereas supervised out-of-domain pre-training yields 23.75 percentage points, revealing a 9.66-point optimistic bias associated with domain overlap. Although out-of-domain pre-training is more realistic in applications where target-domain data are scarce, its effectiveness depends strongly on the compatibility between source and target domains. We further show that labeled source data are not strictly required, with an augmentation-based label-free strategy reaching an average gain of 27.71 percentage points and closely matching supervised out-of-domain pre-training at 27.97 percentage points. Finally, we introduce a descriptor-based source-selection strategy that estimates source-domain suitability before pre-training, reaching a median gap of only 1.37 percentage points to oracle selection. These findings highlight the need to move beyond in-domain pre-training as the default few-shot evaluation protocol, since it can overestimate performance in realistic scenarios where target-domain data are scarce.

Federated pre-training offers a way to train foundation models on private or distributed data without centralizing the underlying datasets. However, evaluating federated pre-training remains challenging because differences in client participation and local data availability can make directly comparable evaluation difficult. Moreover, pre-training test perplexity is tied to the pre-training distribution, while downstream benchmarks introduce task-specific adaptation that may not faithfully reflect the test perplexity established during pre-training. In this work, we study which evaluation protocol more reliably reflects federated pre-training quality. Using a controlled set of centralized and federated-trained models of a 16M parameter transformer model trained on identical client data, we assess evaluation protocols by whether they preserve a reference ranking established on the same pre-training testset. We compare downstream fine-tuning on GLUE, including full, head-only, and reduced-data variants, with next-token prediction on GLUE text as an intrinsic evaluation signal. Our results show that downstream fine-tuning does not reliably preserve the pre-training ranking, whereas direct next-token prediction exhibits a strong correspondence with the pre-training test perplexity. These findings suggest that downstream fine-tuning alone can be misleading when comparing federated pre-trained models, and that evaluation signals closer to the original pre-training objective deserve greater attention.

Real-world model deployments demand strong performance on narrow domains where data is often scarce. Typically, practitioners finetune models to specialize them, but this risks overfitting to the domain and forgetting general knowledge. We study a simple strategy, specialized pretraining (SPT), where a small domain dataset, typically reserved for finetuning, is repeated starting from pretraining as a fraction of the total tokens. Across three specialized domains (ChemPile, MusicPile, and ProofPile), SPT improves domain performance and preserves general capabilities after finetuning compared to standard pretraining. In our experiments, SPT reduces the pretraining tokens needed to reach a given domain performance by up to 1.75x. These gains grow when the target domain is underrepresented in the pretraining corpus: on domains far from web text, a 1B SPT model outperforms a 3B standard pretrained model. Beyond these empirical gains, we derive overfitting scaling laws to guide practitioners in selecting the optimal domain-data repetition for a given pretraining compute budget. Our observations reveal the finetuner's fallacy: while finetuning may appear to be the cheapest path to domain adaptation, introducing specialized domain data during pretraining stretches its utility. SPT yields better specialized domain performance (via reduced overfitting across repeated exposures) and better general domain performance (via reduced forgetting during finetuning), ultimately achieving stronger results with fewer parameters and less total compute when amortized over inference. To get the most out of domain data, incorporate it as early in training as possible.