CSE599J: Data-centric Machine Learning

Winter 2023-2024
Wednesdays and Fridays, 3pm to 4:20pm
CSE2 371
Gradescope | Ed

Pang Wei Koh

Instructor

pangwei@cs.washington.edu

Akari Asai

TA

akari@cs.washington.edu

Office hours by appointment.
To contact course staff, please make an Ed post or email both Pang Wei and Akari.
We welcome feedback on the course. If you prefer to leave it anonymously, use this form.

---

Many advances in machine learning over the past decade have been powered by the increasing availability of larger and more diverse datasets. Where do these datasets come from? What issues are present in these datasets, and how might we deal with them? We will study questions around how we can better use our available data for training, at inference, and for evaluation, as well as ethical, legal, and security issues around data use. The course will be primarily based around paper reading and discussions, with an open-ended course project.

This is a seminar designed for PhD students. Students are expected to be able to read and understand the assigned papers on their own, and they should be familiar with ML and NLP concepts at the level of having taken advanced undergraduate classes.

Schedule

Weekly due dates:

• By Monday 11:59pm: Slides for Wednesday's papers (presenters only)
• By Tuesday 11:59pm: Paper reflections for Wednesday's papers (everyone)
• By Wednesday 11:59pm: Slides for Friday's papers (presenters only)
• By Thursday 11:59pm: Paper reflections for Friday's papers (everyone)

1. Data for pretraining

Jan 3 (Wed)	No class
Jan 5 (Fri)	Course overview & dataset construction (slides) Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus DataComp: In search of the next generation of multimodal datasets Optional reading Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer LAION-5B: An open large-scale dataset for training next generation image-text models Demystifying CLIP Data
Jan 10 (Wed)	Scaling laws (slides) Training Compute-Optimal Large Language Models Scaling Data-Constrained Language Models Optional reading Scaling Laws for Neural Language Models
Jan 12 (Fri)	Data filtering (slides) Deduplicating Training Data Makes Language Models Better Whose Language Counts as High Quality? Measuring Language Ideologies in Text Data Selection Optional reading SemDeDup: Data-efficient learning at web-scale through semantic deduplication Beyond neural scaling laws: beating power law scaling via data pruning Data Filtering Networks Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks Confident Learning: Estimating Uncertainty in Dataset Labels
Jan 17 (Wed)	Dataset composition (slides) A Pretrainer's Guide to Training Data: Measuring the Effects of Data Age, Domain Coverage, Quality, & Toxicity Scaling Laws of Synthetic Images for Model Training ... for Now Optional reading Is a Caption Worth a Thousand Images? A Controlled Study for Representation Learning Quality Not Quantity: On the Interaction between Dataset Design and Robustness of CLIP Data Determines Distributional Robustness in Contrastive Language-Image Pre-training (CLIP) Improving Multimodal Datasets with Image Captioning StableRep: Synthetic Images from Text-to-Image Models Make Strong Visual Representation Learners
Jan 19 (Fri)	Biases in datasets (slides) From Pretraining Data to Language Models to Downstream Tasks: Tracking the Trails of Political Biases Leading to Unfair NLP Models Towards Measuring the Representation of Subjective Global Opinions in Language Models Optional reading Whose Opinions Do Language Models Reflect? Can Large Language Models Capture Dissenting Human Voices? The "Problem" of Human Label Variation: On Ground Truth in Data, Modeling and Evaluation

2. Data for tuning and evaluation

Jan 24 (Wed)	Generative evaluation (slides) Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena All That's 'Human' Is Not Gold: Evaluating Human Evaluation of Generated Text Optional reading LMRL Gym: Benchmarks for Multi-Turn Reinforcement Learning with Language Models Adaptive Testing of Computer Vision Models
Jan 26 (Fri)	Data for alignment (slides) LIMA: Less Is More for Alignment The False Promise of Imitating Proprietary LLMs Optional reading RLAIF: Scaling Reinforcement Learning from Human Feedback with AI Feedback Self-Instruct: Aligning Language Models with Self-Generated Instructions The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning The Flan Collection: Designing Data and Methods for Effective Instruction Tuning How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources
Jan 31 (Wed)	Ambiguity and disagreement (slides) We're Afraid Language Models Aren't Modeling Ambiguity Jury Learning: Integrating Dissenting Voices into Machine Learning Models Optional reading AmbigQA: Answering Ambiguous Open-domain Questions

3. Adapting to different data distributions

Feb 2 (Fri)	Distribution shifts (project proposal due) (slides) WILDS: A benchmark of in-the-wild distribution shifts A theory of learning from different domains Optional reading Extending the WILDS benchmark for unsupervised adaptation Accuracy on the Line: On the Strong Correlation Between Out-of-Distribution and In-Distribution Generalization Out-of-Domain Robustness via Targeted Augmentations
Feb 7 (Wed)	Reweighting data (slides) Improving predictive inference under covariate shift by weighting the log-likelihood function Distributionally Robust Language Modeling Optional reading Distributionally robust neural networks for group shifts DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining Data Selection for Language Models via Importance Resampling
Feb 9 (Fri)	Domain adaptation (slides) Don't Stop Pretraining: Adapt Language Models to Domains and Tasks MEDITRON-70B: Scaling Medical Pretraining for Large Language Models Optional reading Galactica: A Large Language Model for Science The Impact of Large Language Models on Scientific Discovery: a Preliminary Study using GPT-4 Llemma: An Open Language Model for Mathematics OpenWebMath: An Open Dataset of High-Quality Mathematical Web Text Solving Quantitative Reasoning Problems with Language Models

4. Linking model output to training data

Feb 14 (Wed)	Data attribution (slides) Studying Large Language Model Generalization with Influence Functions Simfluence: Modeling the Influence of Individual Training Examples by Simulating Training Runs Optional reading Understanding Black-box Predictions via Influence Functions TRAK: Attributing Model Behavior at Scale
Feb 16 (Fri)	Retrieval-based models (slides) When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories Atlas: Few-shot Learning with Retrieval Augmented Language Models Optional reading Large Language Models Struggle to Learn Long-Tail Knowledge The Web Is Your Oyster - Knowledge-Intensive NLP against a Very Large Web Corpus Improving language models by retrieving from trillions of tokens In-Context Retrieval-Augmented Language Models Generalization through Memorization: Nearest Neighbor Language Models Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks Retrieval-Augmented Multimodal Language Modeling DocPrompting: Generating Code by Retrieving the Docs Dense X Retrieval: What Retrieval Granularity Should We Use? In-Context Pretraining: Language Modeling Beyond Document Boundaries
Feb 21 (Wed)	Memorization (slides) Quantifying Memorization Across Neural Language Models Extracting Training Data from Large Language Models Optional reading Membership Inference Attacks From First Principles Scalable Extraction of Training Data from (Production) Language Models Extracting Training Data from Diffusion Models

5. Legal, ethical, and security considerations

Feb 23 (Fri)	No class
Feb 28 (Wed)	Copyright (slides) Talkin' 'Bout AI Generation: Copyright and the Generative-AI Supply Chain NYT-OpenAI lawsuit Optional reading Foundation Models and Fair Use Glaze: Protecting Artists from Style Mimicry by Text-to-Image Models
Mar 1 (Fri)	Segregating data (slides) What Does it Mean for a Language Model to Preserve Privacy? SILO language models: Isolating legal risk in a nonparametric datastore
Mar 6 (Wed)	Data security and robustness Poisoning Web-Scale Training Datasets is Practical DECODINGTRUST: A Comprehensive Assessment of Trustworthiness in GPT Models Optional reading Poisoning Language Models During Instruction Tuning Universal and Transferable Adversarial Attacks on Aligned Language Models Are aligned neural networks adversarially aligned?
Mar 8 (Fri)	Last class; project presentations
Mar 11 (Mon)	No class; project writeups due

---

Acknowledgements

We are grateful to Sewon Min, Rulin Shao, Ludwig Schmidt, and Tatsunori Hashimoto for their feedback and assistance with this course.