This page is still being polished. If you have thoughts, please share them via the feedback form.
Data on this page is preliminary and may change. Please do not share or cite these figures publicly.
Changes to the model's learned parameters, architecture, or training process, including modifications to training data that affect what the model learns.
Also in AI System
To reduce the toxicity and bias of LLMs, prior efforts mainly focus on enhancing the quality of training data and conducting safety training.
Reasoning
Detoxifying spans both training data curation and safety training objectives, distinct model modification mechanisms.
Toxic and Biased Data Interventions
Similar to the idea of privacy data intervention, toxic/biased data intervention aims to filter undesired content within large-scale web-collected datasets to derive higher-quality training data. For toxicity detection, previous work [247], [248] usually uses labeled datasets to train toxicity classifiers [249]. Some of them have developed advanced automated tools to detect the toxic data in the training corpora, such as Perspective API [250] and Azure AI Content Safety [251]. For data debiasing, the majority of studies [252]–[255] focus on removing or altering bias-related words in the corpora, such as generating a revised dataset by replacing bias-related words (e.g., gendered words) with their opposites [253] or replacing biased texts in the dataset with neutral texts [254]. However, recent work [96] finds that a simple data intervention method may increase LM loss and carry the risk of accidentally filtering out some demographic groups. As a consequence, researchers in LLMs employ varied strategies when addressing toxic and biased data.
1.1.1 Training DataSafety Training
Different from the data intervention-based methods of detoxifying and debiasing, safety training is a training-based method to mitigate toxicity and bias issues. For model detoxifying, several approaches [256]–[258] regard detoxification as a style transfer task, and thus they fine-tune language models to transfer offensive text into non-offensive variants. For model debiasing, a bunch of studies [252], [259]– [262] attempt to use word embedding or adversarial learning to mitigate the impact caused by the proportion gaps between different demographic words. With the development of LLMs, recent works [263], [264] demonstrated that using the training techniques like reinforcement learning from human feedback (RLHF) can effectively improve the performance of detoxifying and debiasing.
1.1.2 Learning ObjectivesMitigation in Input Modules
Mitigating the threat posed by the input module presents a significant challenge for LLM developers due to the diversity of the harmful inputs and adversarial prompts [209], [210].
1.2.1 Guardrails & FilteringMitigation in Input Modules > Defensive Prompt Design
Directly modifying the input prompts is a viable approach to steer the behavior of the model and foster the generation of responsible outputs. This method integrates contextual information or constraints in the prompts to provide background knowledge and guidelines while generating the output [22].
1.2.1 Guardrails & FilteringMitigation in Input Modules > Malicious Prompt Detection
Different from the methods of designing defensive prompts to preprocess the input, the malicious prompt detection method aims to detect and filter out the harmful prompts through the input safeguard
1.2.1 Guardrails & FilteringMitigation in Language Models
This section delves into mitigating risks associated with models, encompassing privacy preservation, detoxification and debiasing, mitigation of hallucinations, and defenses against model attacks.
1.1 ModelMitigation in Language Models > Privacy Preserving
Privacy leakage is a crucial risk of LLMs, since the powerful memorization and association capabilities of LLMs raise the risk of revealing private information within the training data. Researchers are devoted to designing privacypreserving frameworks in LLMs [226], [227], aiming to safeguard sensitive PII from possible disclosure during humanmachine conservation
1.1 ModelMitigation in Language Models > Hallucination Mitigation
Hallucinations, one of the key challenges associated with LLMs, have received extensive studies.
1.2.9 OtherRisk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems
Cui, Tianyu; Wang, Yanling; Fu, Chuanpu; Xiao, Yong; Li, Sijia; Deng, Xinhao; Liu, Yunpeng; Zhang, Qinglin; Qiu, Ziyi; Li, Peiyang; Tan, Zhixing; Xiong, Junwu; Kong, Xinyu; Wen, Zujie; Xu, Ke; Li, Qi (2024)
Despite their impressive capabilities, large lan- guage models (LLMs) have been observed to generate responses that include inaccurate or fabricated information, a phenomenon com- monly known as “hallucination”. In this work, we propose a simple Induce-then-Contrast De- coding (ICD) strategy to alleviate hallucina- tions. We first construct a factually weak LLM by inducing hallucinations from the original LLMs. Then, we penalize these induced hallu- cinations during decoding to enhance the fac- tuality of the generated content. Concretely, we determine the final next-token predictions by amplifying the predictions from the orig- inal model and downplaying the induced un- truthful predictions via contrastive decoding. Experimental results on both discrimination- based and generation-based hallucination eval- uation benchmarks, such as TruthfulQA and FACTSCORE, demonstrate that our proposed ICD methods can effectively enhance the factu- ality of LLMs across various model sizes and families. For example, when equipped with ICD, Llama2-7B-Chat and Mistral-7B-Instruct achieve performance comparable to ChatGPT and GPT4 on TruthfulQA, respectively.
Build and Use Model
Training, fine-tuning, and integrating the AI model
Developer
Entity that creates, trains, or modifies the AI system
Manage
Prioritising, responding to, and mitigating AI risks
Primary
1 Discrimination & Toxicity