Reading List Model Lakes

Formally defines a model lake as a repository for heterogeneous ML models — by analogy to data lakes — and lays out the research agenda. Identifies four core model-lake tasks (attribution, versioning, search, benchmarking) and argues that today's "20th-century keyword search over manually specified names or metadata" is inadequate for repositories with millions of fine-tuned variants of foundation models. Draws explicit lessons from data lake research to motivate the model lake research program.

A complementary vision focused on enterprise model lifecycle management, audit, and governance. Extends the model-lake concept to integrate datasets, code, and models in a single ecosystem with operational benefits for versioning, discovery, and reuse. Useful as a counterpoint to Pal et al.'s research-oriented framing.

An early end-to-end model management system supporting experiment tracking, versioning, comparison, and reproducibility through a model store with native clients for scikit-learn and Spark MLlib. Establishes the basic primitives — log models with parameters, pipelines, and metadata; query and compare via a frontend — that every subsequent model management system has built on.

A position paper surveying the conceptual, engineering, and data-management challenges of managing ML models in production. Catalogues the open problems — model versioning, prediction validation, retraining decisions, pipeline lineage — that the field has been working on since. Required background for understanding why model lakes are hard.

Introduces MLflow, now the most widely deployed open-source ML platform: tracking (logging runs, parameters, metrics), projects (reproducible packaging), models (a standard format for downstream tools), and the model registry (centralized model store with lineage, versioning, stage transitions). De facto reference architecture for production model management.

A provenance-centric system for collaborative data-science workflows. Tracks the full graph of datasets, transformations, models, and analyses across users and over time, supporting queries like "which models depend on this dataset version" — a capability that becomes essential at model-lake scale.

A systematic study of how modern ML platforms (MLflow, others) capture provenance across the full pipeline lifecycle — registered models, version creation, deletions, stage transitions. Useful for understanding both what current systems track and what they miss for model-lake-scale governance.

Proposes "model cards" — short structured documents accompanying released models — covering intended use, performance across demographic groups, training data, and ethical considerations. Now the standard model-documentation framework on Hugging Face and most major model hubs, and the lever by which much of model-lake metadata becomes available at all.

A large-scale empirical analysis of 32,000 model cards on Hugging Face. Shows that even where cards exist, they're frequently incomplete, inconsistent, or missing critical sections — revealing the gap between Mitchell et al.'s ideal and current practice, and motivating automated model-card generation.

Uses LLMs to automatically generate model and data cards from training-time metadata and code artifacts. A concrete step toward the model-lake vision of metadata at scale: when the lake has millions of models, hand-curated cards are no longer feasible.

Builds a structured "atlas" of the Hugging Face model ecosystem by recovering the implicit fine-tuning DAG that connects foundation models to their descendants. Demonstrates concrete applications: predicting model attributes, recovering fine-tuning structure even when undocumented, and surfacing the evolutionary structure of public model hubs. A model-lake-style approach applied to today's largest public model repository.

Recasts model selection as an iterative reasoning process — Reasoning, Retrieval, Refinement, Reflection — rather than one-shot retrieval. Targets the 2M+ model setting of Hugging Face explicitly, where putting full descriptions into prompts is infeasible. Mirrors the agentic shift now happening in text-to-SQL schema linking.

A broader framing covering not just models but the full bundle of curated ML assets — models, datasets, prompts, agents. Useful for situating model search within the wider problem of asset discovery in the LLM era.

A foundational empirical analysis of how the Hugging Face model hub is actually used: how models are named, packaged, reused, and forked. Documents the patterns and pathologies (inconsistent naming, missing documentation, broken dependencies) that any production model lake must contend with.

A qualitative and empirical analysis of model naming practices on Hugging Face. Shows that names are inconsistent, sometimes misleading, and frequently inadequate as a basis for selection — a direct argument for the richer search and retrieval primitives that model lake research targets.