三小时掌握大型语言模型的公平性

This tutorial is grounded in oursurveys and establishedbenchmarks,all available as open-sourceresources:https://github.com/LavinWong/Fairness-in-Large-Language-Model WARNING: The following slides contains examples of model bias andevaluation which are offensive in nature. Large Language Models are fascinating! Diverse ApplicationsAcross Industries Breaking Language andKnowledge Boundaries Unprecedented LanguageCapabilities But they are not perfect! LLMs exhibit unfairness intheir answers! But they are not perfect! LLMs exhibit unfairness intheir answers! Emergency need to handle bias inLLMs’ behavior! Bias mitigating in LLMs is different How bias is formed INLARGELANGUAGEMODELS How to measure unfairness What methods can be applied to mitigate bias What are the tools for measuring and mitigating bias Why is mitigating bias challenged Bias mitigating in LLMs is different How bias is formed INLARGELANGUAGEMODELS How to measure unfairness What methods can be applied to mitigate bias What are the tools for measuring and mitigating bias Why is mitigating bias challenged We built a roadmap to explore these questions! Roadmap Section 1: Background on LLMs Section 2: Quantifying bias in LLMs Section 3: Mitigating bias in LLMs Section 4: Resources for evaluating bias in LLMs Section 5: Challenges and future directions Section 1: Background on LLMs Content ➢Review the development history ofLLMs ➢Training procedure of LLMs, how itachieve such capabilities ➢Explore the bias sources in LLMs 1.1 History of LLMs This section is grounded in ourintroduction to LLMs survey [1]. [1] Wang, Zichong, Chu, Zhibo, Doan, Thang Viet, Ni, Shiwen, Yang, Min, Zhang, Wenbin.“History,development, and principles of large language models: an introductory survey."AI and Ethics(2024): 1-17. 1.1 History of LLMs a. Language Models ●Earlier Stages:Statistical LMs -> Neural LMs●N-grams [2]: ●For example: [2] Jurafsky, Dan; Martin, James H. (7 January 2023). "N-gram Language Models". Speech and LanguageProcessing (PDF) (3rd edition drafted.). Retrieved 24 May 2022. 1.1 History of LLMs a. Language Models ●Earlier Stages:Statistical LMs -> Neural LMs●Word2Vec [3,4]: 14[3] Mikolov T, Chen K, Corrado G, Dean J (2013) Efficient estimation of word representations in vector space. In:Proceedings of ICLR Workshop 2013[4] Mikolov T, Sutskever I, Chen K, Corrado G, Dean J (2013) Distributed representations of words and phrases andtheir compositionality. Adv Neural Inf Process Syst 26:1 1.1 History of LLMs a. Language Models ●Earlier Stages:Statistical LMs -> Neural LMs●RNN [5]: [5] A. Graves, A. -r. Mohamed and G. Hinton, "Speech recognition with deep recurrent neural networks," 2013IEEE International Conference on Acoustics, Speech and Signal Processing, Vancouver, BC, Canada, 2013, pp.6645-6649, doi: 10.1109/ICASSP.2013.6638947. 1.1 History of LLMs a. Language Models ●Drawbacks: ○Poor generalization○Lack of long-termdependence○Recurrent computation○Difficult in capturingcomplex linguisticproperties and phenomena 1.1 History of LLMs Until Transformers [6] … 1.1 History of LLMs b. Large Language Models ●Until Transformers:○Self-Attention:Long-Range Dependencies 1.1 History of LLMs b. Large Language Models ●Until Transformers:○Multi-head Attention:Contextualized Word Representations 1.1 History of LLMs b. Large Language Models ●Until Transformers:○Parallelization andScalability 1.1 History of LLMs b. Large Language Models ●Transformers revolutionized the natural languageprocessing landscape!●Results in a massive blooming era of LLMs: GPT,BERT, LLaMA, Claude and more to go!●Broad applications across domains:○Education○Healthcare○Technology○And so on… 1.2 Training LLMs Key steps to train LLMs ●Training large language models is a complex,multi-step process that requires careful planning andexecution. 1.2 Training LLMs a. Data Preparation ●Data is the foundation of LLMs.●“Garbage In, Garbage Out”:Poor data quality can lead to biased,inaccurate, or unreliable model outputs.●High-quality data can lead to accurate,coherent, and reliable outputs. [7] Srivastava, Ankit, Piyush Makhija, and Anuj Gupta. "Noisy Text Data: Achilles’ Heel of BERT." Proceedings ofthe Sixth Workshop on Noisy User-generated Text (W-NUT 2020). 2020. 1.2 Training LLMs a. Data Preparation●Quality: Accurately represent the domain and language style, factually correct and free from errors. ●Examples: 1.2 Training LLMs a. Data Preparation ●Diversity: Represent a wide variety oflanguages, domains, and contexts to improvegeneralization.●Some languages have limited availability of linguistic data, tools, and resources comparedto more widely spoken languages. 1.2 Training LLMs a. Data Preparation ●Data Cleaning - Quality Filtering: ○Noise/Outlier Handling: Identifying and removing noisy orirrelevant data that could distort the model’s performance.○Normalization: Ensuring that the data is consistent andstandardized