We introduce NNsight and NDIF, technologies that work in tandem to enable scientific study of the representations and computations learned by very large neural networks. NNsight is an open-source system that extends PyTorch to introduce deferred remote execution. The National Deep Inference Fabric (NDIF) is a scalable inference service that executes NNsight requests, allowing users to share GPU resources and pretrained models. These technologies are enabled by the Intervention Graph, an architecture developed to decouple experimental design from model runtime. Together, this framework provides transparent and efficient access to the internals of deep neural networks such as very large language models (LLMs) without imposing the cost or complexity of hosting customized models individually. We conduct a quantitative survey of the machine learning literature that reveals a growing gap in the study of the internals of large-scale AI. We demonstrate the design and use of our framework to address this gap by enabling a range of research methods on huge models. Finally, we conduct benchmarks to compare performance with previous approaches. Code, documentation, and tutorials are available at https://nnsight.net/.

Much of the knowledge encoded in transformer language models (LMs) may be expressed in terms of relations: relations between words and their synonyms, entities and their attributes, etc. We show that, for a subset of relations, this computation is well-approximated by a single linear transformation on the subject representation. Linear relation representations may be obtained by constructing a first-order approximation to the LM from a single prompt, and they exist for a variety of factual, commonsense, and linguistic relations. However, we also identify many cases in which LM predictions capture relational knowledge accurately, but this knowledge is not linearly encoded in their representations. Our results thus reveal a simple, interpretable, but heterogeneously deployed knowledge representation strategy in transformer LMs.

We report the presence of a simple neural mechanism that represents an input-output function as a vector within autoregressive transformer language models (LMs). Using causal mediation analysis on a diverse range of in-context-learning (ICL) tasks, we find that a small number attention heads transport a compact representation of the demonstrated task, which we call a function vector (FV). FVs are robust to changes in context, i.e., they trigger execution of the task on inputs such as zero-shot and natural text settings that do not resemble the ICL contexts from which they are collected. We test FVs across a range of tasks, models, and layers and find strong causal effects across settings in middle layers. We investigate the internal structure of FVs and find while that they often contain information that encodes the output space of the function, this information alone is not sufficient to reconstruct an FV. Finally, we test semantic vector composition in FVs, and find that to some extent they can be summed to create vectors that trigger new complex tasks. Our findings show that compact, causal internal vector representations of function abstractions can be explicitly extracted from LLMs.

Social media platforms and online streaming services have spawned a new breed of Hate Speech (HS). Due to the massive amount of user-generated content on these sites, modern machine learning techniques are found to be feasible and cost-effective to tackle this problem. However, linguistically diverse datasets covering different social contexts in which offensive language is typically used are required to train generalizable models. In this paper, we identify the shortcomings of existing Bangla HS datasets and introduce a large manually labeled dataset BD-SHS that includes HS in different social contexts. The labeling criteria were prepared following a hierarchical annotation process, which is the first of its kind in Bangla HS to the best of our knowledge. The dataset includes more than 50,200 offensive comments crawled from online social networking sites and is at least 60% larger than any existing Bangla HS datasets. We present the benchmark result of our dataset by training different NLP models resulting in the best one achieving an F1-score of 91.0%. In our experiments, we found that a word embedding trained exclusively using 1.47 million comments from social media and streaming sites consistently resulted in better modeling of HS detection in comparison to other pre-trained embeddings. Our dataset and all accompanying codes is publicly available at github.com/naurosromim/hate-speech-dataset-for-Bengali-social-media