DAG Cards - AI Developer Tools Tool

Overview

DAG Cards is an open-source, template-driven generator that builds static HTML “DAG cards” from Metaflow Flow classes. The tool combines structural DAG metadata (steps, dependencies, docstrings) with runtime information (recent runs, metrics, artifacts) pulled from experiment tracking services to produce a Google-style one-page report for an ML pipeline. The project ships a Jinja HTML template and a small builder script (card_builder.py) that assembles JSON from Metaflow and Weights & Biases (W&B), renders charts with Chart.js, and writes a self-contained HTML page into a card/ folder for distribution and review. ([github.com](https://github.com/jacopotagliabue/dag-card-is-the-new-model-card)) DAG Cards was developed as part of a data-centric-AI discussion (paper accepted at the DCAI workshop, NeurIPS 2021) and published alongside a short paper describing the motivation for documenting pipelines rather than single models. The repository’s README lists the card structure (Overview, Owners, DAG, Model, Tests) and explains the expected inputs and runtime dependencies (Metaflow, W&B, Jinja). The implementation targets Python 3.6-era environments and is intended as an easily-extensible MVP rather than a polished UI library. ([arxiv.org](https://arxiv.org/abs/2110.13601?utm_source=openai))

GitHub Statistics

• Stars: 30
• Forks: 7
• Contributors: 1
• License: MIT
• Primary Language: Python
• Last Updated: 2021-12-07T13:29:22Z

Repository activity indicates a small, single-maintainer project with modest community traction: 30 stars and 7 forks on GitHub, no open issues or pull requests, and a short commit history (22 commits) with the last commit dated December 7, 2021. The project is licensed under the MIT license and contains a requirements.txt enumerating exact dependency versions (metaflow, wandb, tensorflow, keras, jinja2, python-dotenv). Overall, the repo serves as a working prototype and research artifact rather than a production-grade, actively-maintained package. ([github.com](https://github.com/jacopotagliabue/dag-card-is-the-new-model-card))

Installation

Install via pip:

git clone https://github.com/jacopotagliabue/dag-card-is-the-new-model-card.git

cd dag-card-is-the-new-model-card

python3 -m venv .venv && source .venv/bin/activate

pip install -r requirements.txt

cp .env.template .env && edit .env to add W&B and Metaflow credentials

METAFLOW_PROFILE=my_profile python card_builder.py  # produces static HTML in card/ folder

METAFLOW_PROFILE=my_profile python training_flow.py run  # optional: run sample DAG to generate W&B runs

Key Features

• Generate static, shareable HTML DAG cards from Metaflow Flow classes using Jinja templates.
• Integrates with Weights & Biases to surface run metrics, artifacts, and latest-K run summaries.
• Renders a visual DAG diagram plus collapsible step-level details and step metadata.
• Includes Owners section and run-distribution charts to show who ran or owns the pipeline.
• Reports quantitative tests and behavioral test outputs alongside model artifacts and paths.

Community

Community engagement is limited: the repo is a small research/engineering artifact with 30 stars, 7 forks, and low ongoing activity (no open issues or PRs). The idea has academic visibility—paper accepted at DCAI@NeurIPS 2021 and discussed in a Towards Data Science blog—indicating interest in the concept even if the codebase itself remains a lightweight prototype. Prospective users should expect to adapt the template and builder script for their environment. ([github.com](https://github.com/jacopotagliabue/dag-card-is-the-new-model-card))

Key Information

• Category: Developer Tools
• Type: AI Developer Tools Tool