Detectors for Safe and Reliable LLMs: Implementations, Uses, and Limitations

Large language models (LLMs) are susceptible to a variety of risks, from non-faithful output to biased and toxic generations. Due to several limiting factors surrounding LLMs (training cost, API access, data availability, etc.), it may not always be feasible to impose direct safety constraints on a deployed model. Therefore, an efficient and reliable alternative is required. To this end, we present our ongoing efforts to create and deploy a library of detectors: compact and easy-to-build classification models that provide labels for various harms. In addition to the detectors themselves, we discuss a wide range of uses for these detector models - from acting as guardrails to enabling effective AI governance. We also deep dive into inherent challenges in their development and discuss future work aimed at making the detectors more reliable and broadening their scope.

Efficient Models for the Detection of Hate, Abuse and Profanity

Large Language Models (LLMs) are the cornerstone for many Natural Language Processing (NLP) tasks like sentiment analysis, document classification, named entity recognition, question answering, summarization, etc. LLMs are often trained on data which originates from the web. This data is prone to having content with Hate, Abuse and Profanity (HAP). For a detailed definition of HAP, please refer to the Appendix. Due to the LLMs being exposed to HAP content during training, the models learn it and may then generate hateful or profane content. For example, when the open-source RoBERTa model (specifically, the RoBERTA base model) from the HuggingFace (HF) Transformers library is prompted to replace the mask token in `I do not know that Persian people are that MASK` it returns the word `stupid` with the highest score. This is unacceptable in civil discourse.The detection of Hate, Abuse and Profanity in text is a vital component of creating civil and unbiased LLMs, which is needed not only for English, but for all languages. In this article, we briefly describe the creation of HAP detectors and various ways of using them to make models civil and acceptable in the output they generate.

An Empirical Investigation into the Effect of Parameter Choices in Knowledge Distillation

We present a large-scale empirical study of how choices of configuration parameters affect performance in knowledge distillation (KD). An example of such a KD parameter is the measure of distance between the predictions of the teacher and the student, common choices for which include the mean squared error (MSE) and the KL-divergence. Although scattered efforts have been made to understand the differences between such options, the KD literature still lacks a systematic study on their general effect on student performance. We take an empirical approach to this question in this paper, seeking to find out the extent to which such choices influence student performance across 13 datasets from 4 NLP tasks and 3 student sizes. We quantify the cost of making sub-optimal choices and identify a single configuration that performs well across the board.

Muted: Multilingual Targeted Offensive Speech Identification and Visualization

Offensive language such as hate, abuse, and profanity (HAP) occurs in various content on the web. While previous work has mostly dealt with sentence level annotations, there have been a few recent attempts to identify offensive spans as well. We build upon this work and introduce Muted, a system to identify multilingual HAP content by displaying offensive arguments and their targets using heat maps to indicate their intensity. Muted can leverage any transformer-based HAP-classification model and its attention mechanism out-of-the-box to identify toxic spans, without further fine-tuning. In addition, we use the spaCy library to identify the specific targets and arguments for the words predicted by the attention heatmaps. We present the model's performance on identifying offensive spans and their targets in existing datasets and present new annotations on German text. Finally, we demonstrate our proposed visualization tool on multilingual inputs.

A Comparative Analysis of Task-Agnostic Distillation Methods for Compressing Transformer Language Models

Large language models have become a vital component in modern NLP, achieving state of the art performance in a variety of tasks. However, they are often inefficient for real-world deployment due to their expensive inference costs. Knowledge distillation is a promising technique to improve their efficiency while retaining most of their effectiveness. In this paper, we reproduce, compare and analyze several representative methods for task-agnostic (general-purpose) distillation of Transformer language models. Our target of study includes Output Distribution (OD) transfer, Hidden State (HS) transfer with various layer mapping strategies, and Multi-Head Attention (MHA) transfer based on MiniLMv2. Through our extensive experiments, we study the effectiveness of each method for various student architectures in both monolingual (English) and multilingual settings. Overall, we show that MHA transfer based on MiniLMv2 is generally the best option for distillation and explain the potential reasons behind its success. Moreover, we show that HS transfer remains as a competitive baseline, especially under a sophisticated layer mapping strategy, while OD transfer consistently lags behind other approaches. Findings from this study helped us deploy efficient yet effective student models for latency-critical applications.

Neural Architecture Search for Effective Teacher-Student Knowledge Transfer in Language Models

Large pre-trained language models have achieved state-of-the-art results on a variety of downstream tasks. Knowledge Distillation (KD) of a smaller student model addresses their inefficiency, allowing for deployment in resource-constraint environments. KD however remains ineffective, as the student is manually selected from a set of existing options already pre-trained on large corpora, a sub-optimal choice within the space of all possible student architectures. This paper proposes KD-NAS, the use of Neural Architecture Search (NAS) guided by the Knowledge Distillation process to find the optimal student model for distillation from a teacher, for a given natural language task. In each episode of the search process, a NAS controller predicts a reward based on a combination of accuracy on the downstream task and latency of inference. The top candidate architectures are then distilled from the teacher on a small proxy set. Finally the architecture(s) with the highest reward is selected, and distilled on the full downstream task training set. When distilling on the MNLI task, our KD-NAS model produces a 2 point improvement in accuracy on GLUE tasks with equivalent GPU latency with respect to a hand-crafted student architecture available in the literature. Using Knowledge Distillation, this model also achieves a 1.4x speedup in GPU Latency (3.2x speedup on CPU) with respect to a BERT-Base Teacher, while maintaining 97% performance on GLUE Tasks (without CoLA). We also obtain an architecture with equivalent performance as the hand-crafted student model on the GLUE benchmark, but with a 15% speedup in GPU latency (20% speedup in CPU latency) and 0.8 times the number of parameters