Model Maps: A Documentation Framework for Tracing Foundation Model Ancestry and Understanding AI Entanglement
Organizations: University of Washington, USA · Carnegie Mellon University, USA · Amazon · Georgetown University, USA
Abstract
The development of foundation models incorporates pre-existing AI artifacts like models, datasets, and benchmarks. These relationships, as part of the broader AI supply chain, produce "entangled" webs in which properties of artifacts upstream can influence the outputs of a foundation model downstream - a phenomenon known as AI entanglement. Without visibility into these connections, a black-box model can inherit upstream properties that result in security vulnerabilities, hallucinations, or biased predictions, yet developers are left with no recourse to identify or mitigate the root issue. To address these concerns, we introduce model maps as a conceptual documentation framework for tracing model ancestry that builds on previous documentation frameworks like ecosystem graphs, datasheets, and model cards. A model map for a given foundation model is a detailed knowledge graph that represents the relationships between artifacts that influence the resulting foundation model. Our framework incorporates a taxonomy of 13 dependency mechanisms including model fine-tuning and synthetic data generation. We apply this framework to construct model maps of 15 closed and open-weight foundation models to illustrate the complexity of AI entanglement in current model development practices. From our model maps, we record 2,778 unique dependencies and trace artifacts that are located up to 9 hops away from a resulting foundation model. Our findings suggest that AI transparency requirements should extend beyond individual models and training-data disclosures to encompass upstream dependencies, which would enable more effective risk assessment, auditing, and accountability across the AI supply chain. To help developers, regulators, and researchers understand and analyze the interactions that can influence the behavior of a foundation model, we release these these model maps at https://modelmaps.ai.
Figures & tables
| Type | Dependency | Source | Target | Description (source A influences target B) |
| Data-based | Data composition | Data | Data | B is composed of data from A |
| Training data | Data | Model | B is trained on data from A | |
| Benchmark evaluation | Data | Model | B is evaluated on data from A | |
| Model-based data curation | Data generation | Model | Data | B incorporates data generated from A |
| Data annotation | Model | Data | B incorporates data annotated from A | |
| Data filtering | Model | Data | B is filtered through scores or predictions from A |
| Model | Alpaca ( Taori et al., 2023 ) | GPT-4 ( OpenAI et al., 2024 ) | Mistral 7B ( Jiang et al., 2023 ) | Apple FM ( Gunter et al., 2024 ) | Llama 3 ( Grattafiori et al., 2024 ) | Phi 4 ( Abdin et al., 2024 ) | DeepSeek-V3 ( DeepSeek-AI et al., 2025 ) | Gemma 3 ( Team et al., 2025 ) | Qwen3 ( Bai et al., 2023 ) | Claude 4 ( Anthropic, 2025 ) | Gemini 2.5 ( Comanici et al., 2025 ) | SmolLM3 ( Bakouch et al., 2025 ) | DR Tulu ( Shao et al., 2026 ) | OLMo 3 ( Olmo et al., 2025 ) | Nemotron 3 ( NVIDIA et al., 2025 ) |
| Nodes | 6 | 38 | 31 | 71 | 121 | 75 | 167 | 46 | 90 | 56 | 115 | 418 | 104 | 681 | 259 |
| Root nodes | 2 | 10 | 2 | 13 | 21 | 14 | 5 | 12 | 11 | 12 | 11 | 4 | 5 | 42 | 13 |
| Datasets | 1 | 6 | 3 | 12 | 17 | 9 | 32 | 3 | 6 | 9 | 13 | 168 | 26 | 270 | 67 |
| Benchmarks | 1 | 15 | 21 | 30 | 59 | 37 | 66 | 15 | 45 | 23 | 50 | 111 | 28 | 166 | 78 |
| Websites | 0 | 4 | 3 | 16 | 21 | 12 | 30 | 6 | 9 | 2 | 14 | 56 | 17 | 71 | 38 |
| Models | 3 | 7 | 4 | 12 | 22 | 15 | 35 | 19 | 29 | 18 | 30 | 75 | 30 | 163 | 71 |
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
| Category | Code | Description ("We form this dependency to…") |
| Data quality | Improve downstream quality | Ensure downstream data is higher quality according to the desired criteria, such as filtering relevant content or isolating to a single language |
| Select a quality upstream source | Choose a data source that is quality (e.g. desirable, appropriate, or relevant) for the task at hand | |
| Efficiency | Increase scale | Expand the data without an explicit purpose of quality or diversity; solely to increase size of the dataset |
| Increase speed | Select an upstream source that is fast or time-saving | |
| Save resources | Decrease costs or use of compute resources | |
| Increase efficiency | Improve efficiency without specifying in which manner |
| Data quality | Example excerpts |
| Improve downstream quality | “…we design a model-based filter by iteratively developing a prompt to use with Llama-3.1-70B…ensuring that we remove as much undesirable data as possible.” ( Albalak et al., 2025 ) |
| “We use Llama3.3-70B-Instruct to filter out these defective Q-A pairs.” ( Fan et al., 2025 ) | |
| Select a quality upstream source | “AutoMathText [is] drawn from multiple high-value sources (e.g., OpenWebMath, arXiv, Algebraic Stack).” ( Zhang et al., 2025 ) |
| “We define two criteria for our training datasets: 1) the samples must be informative and well-structured…” ( Messmer et al., 2025 ) |
| Efficiency | Example excerpts |
| Increase scale | “Unfortunately, such large scale prevents human annotations of toxicity (e.g., we score at least 80 GB of text in §6). Therefore, we rely on Perspective API…” ( Gehman et al., 2020 ) |
| Increase speed | “By requiring only a small set of few-shot examples, we eliminated the need for extensive human annotation, facilitating rapid policy adaptation…” ( Zeng et al., 2025 ) |
| Save resources | “To further enhance data processing efficiency and reduce hardware resource requirements, in this paper we also develop a text evaluation model…” ( Chen et al., 2023a ) |
| Increase efficiency | “We use DistilRoberta to generate quality scores for each document for efficiency reasons.” ( Grattafiori et al., 2024 ) |
| Generalization | Example excerpts |
| Represent the real world | “By fine-tuning the model on high-quality code from GitHub, we ensured it was exposed to real-world examples…” ( Huang et al., 2025b ) |
| Include diverse data | “We choose the YT-Temporal-1Bn dataset as the source for Neptune, because of its large and diverse corpus…” ( Nagrani et al., 2024 ) |
| Select a versatile upstream model | “The LLM serves as a versatile processor capable of transforming unstructured information from diverse scenarios into rewards…” ( DeepSeek-AI et al., 2025 ) |
| Improve downstream model generalization | “We incorporated multilingual datasets to ensure that our model could handle a wide range of languages” ( Abdin et al., 2024 ) |
| Model performance | Example excerpts |
| Improve downstream performance | “To teach Nemotron 3 Nano to complete autonomous terminal-based tasks, we generate a diverse set of verifiable tasks based on Terminal Bench.” ( NVIDIA et al., 2025 ) |
| “Microanneals show that inclusion of this data results in substantial improvements to math and coding tasks…” ( Olmo et al., 2025 ) | |
| Select a performant upstream source | “We used an external model to perform the translation, instead of relying on GPT-4 itself, in case the model had unrepresentative performance for its own translations.” ( Grattafiori et al., 2024 ) |
| “We initially utilize GPT-4’s question-rewriting capabilities to generate question variations…” ( Li et al., 2024c ) |
| Novelty | Example excerpts |
| Overcome an existing gap | “The primary challenge we encountered when building the reasoning mode dataset was the scarcity of datasets containing reasoning traces for certain domains. To address this gap, we generated synthetic data by prompting Qwen3-32B….” ( Bakouch et al., 2025 ) |
| “To mitigate the impact of outdated programming syntax and obsolete external library interfaces in the pre-training dataset, we synthesized a tool usage instruction tuning dataset using up-to-date external library documentation.” ( Huang et al., 2025b ) | |
| Differentiate from prior work | “Compared with Qwen2.5, we have significantly expanded the scale and diversity of our training data.” ( Bai et al., 2023 ) |
| “We employed a methodology similar to that of WizardCoder, with the exception that ours is open-source.” ( Luo et al., 2024 ) |
| Prior work | Example excerpts |
| Follow prior process | “As with prior GPT models, we fine-tune the model’s behavior using reinforcement learning with human feedback” ( OpenAI et al., 2024 ) |
| “We adhere to the original configurations of AlpacaEval 2.0 and Arena-Hard, which leverage GPT-4-Turbo-1106 as judges…” ( DeepSeek-AI et al., 2025 ) | |
| Select a commonly used source | “In this work, we directly adopt starcoderdata as our seed corpus, a filtered version of The Stack (Kocetkov et al., 2022) dataset that StarCoder is trained on…We chose starcoderdata because it is widely adopted…” ( Wei et al., 2023 ) |
| “While EvalPlus is general, this paper focuses on evaluating its effectiveness on HumanEval, one of the most widely-used datasets for code generation” ( Liu et al., 2023 ) |
| Research standards | Example excerpts |
| Select an easily accessed source | “To prepare our email subject line dataset, we use the Enron dataset which is a collection of email messages of employees in the Enron Corporation. We use Enron because it can be released to the public…” ( Zhang and Tetreault, 2019 ) |
| “Our considerations in selecting a corpus of documents for which to collect summaries are …(3) The documents have a permissive license such that they can be easily distributed” ( Wang et al., 2022 ) | |
| Decontaminate | “To thoroughly prevent data leakage from the test datasets to the training dataset, we implemented an additional data filtering step. Utilizing the SOTA embeddings model, gte-large…” ( Luo et al., 2024 ) |
| “To implement rigorous benchmark decontamination, we follow the approach of Toshniwal et al. and He et al. by deploying LLM-based decontamination…” ( Fan et al., 2025 ) |
| Safety | Example excerpts |
| Filter sensitive content | “All toxic conversations identified by the OpenAI Moderations API or Detoxify have been removed from this version of the dataset.” ( Zhao et al., 2024 ) |
| “This section outlines our efforts to remove Personally Identifiable Information (PII) from the training data.” ( Li et al., 2023 ) | |
| Measure sensitive content | “We used an automatic grading system (auto-grader), designed to evaluate model outputs by policy, to determine the harmfulness of the model’s response.” ( Anthropic, 2025 ) |
| “The evaluation reports the safety of outputs using the WildGuard classifier’s safety label…” ( Han et al., 2024 ) | |
| Improve safety | “We use a mix of reviewers and automated systems to identify and enforce against misuse of our models. Our automated systems include a suite of machine learning and rule-based classifier detections that identify content that might violate our policies.” ( OpenAI et al., 2024 ) |
| “To mitigate these hallucinations, we introduced a targeted dataset into Claude Opus 4’s training…” ( Anthropic, 2025 ) |