TabulaSage
An open-source offline data analysis tool that uses scikit-learn's IsolationForest and decision tree algorithms. It supports column profiling, anomaly detection, and explainable predictive analysis on CSV files. All computations are performed locally, with no data leaving the user's machine.
Timeline (last 90 days)
TabulaSage is an offline data analyzer built by developer Zhiyilang for the Hacktoberfest 2026 open-source AI challenge, using open-source AI as its entire engine and ensuring files never leave the local machine.
TabulaSage provides three types of analysis for CSV files: per-column profiling (type, missing rate, distribution, high-frequency categories), anomaly detection, and target-column driving factors produced by an interpretable model.
There are no network calls anywhere in TabulaSage's entire code path, and the code is approximately 200 lines of Python.
TabulaSage uses scikit-learn's IsolationForest for anomaly detection and a shallow decision tree (DecisionTreeClassifier) as the interpretable model.
In a 15-row sample with one row of bad data injected, IsolationForest detected the anomalous data in row 9 (age=999, income=15).
In sample testing, the interpretable model achieved 100% test accuracy, with the strongest driving factor being age (importance 0.57), followed by monthly_spend.
TabulaSage uses 'mean + 2σ' as a parameter-free anomaly flagging threshold rather than relying on scikit-learn's default contamination parameter, in order to avoid falsely flagging approximately 40% of data on small datasets.
In the first version of TabulaSage, failing to remove the --target column from the features caused label leakage, producing a spurious 100% accuracy; the fix was to remove the target column from the feature list.
TabulaSage can integrate a locally running open-weight LLM (such as Gemma running via Ollama) to generate richer descriptions, or integrate TabPFN for predictions, without rewriting the application.
TabulaSage's insight generation uses deterministic templates to transform statistical data into sentences, requiring no LLM.
1 more timeline events
All Facts (11)
TabulaSage is an offline data analyzer built by developer Zhiyilang for the Hacktoberfest 2026 open-source AI challenge, using open-source AI as its entire engine and ensuring files never leave the local machine.
50%UnverifiedTabulaSage provides three types of analysis for CSV files: per-column profiling (type, missing rate, distribution, high-frequency categories), anomaly detection, and target-column driving factors produced by an interpretable model.
50%UnverifiedThere are no network calls anywhere in TabulaSage's entire code path, and the code is approximately 200 lines of Python.
50%UnverifiedTabulaSage uses scikit-learn's IsolationForest for anomaly detection and a shallow decision tree (DecisionTreeClassifier) as the interpretable model.
50%UnverifiedIn a 15-row sample with one row of bad data injected, IsolationForest detected the anomalous data in row 9 (age=999, income=15).
50%UnverifiedIn sample testing, the interpretable model achieved 100% test accuracy, with the strongest driving factor being age (importance 0.57), followed by monthly_spend.
50%UnverifiedTabulaSage uses 'mean + 2σ' as a parameter-free anomaly flagging threshold rather than relying on scikit-learn's default contamination parameter, in order to avoid falsely flagging approximately 40% of data on small datasets.
50%UnverifiedIn the first version of TabulaSage, failing to remove the --target column from the features caused label leakage, producing a spurious 100% accuracy; the fix was to remove the target column from the feature list.
50%UnverifiedTabulaSage can integrate a locally running open-weight LLM (such as Gemma running via Ollama) to generate richer descriptions, or integrate TabPFN for predictions, without rewriting the application.
50%UnverifiedTabulaSage's insight generation uses deterministic templates to transform statistical data into sentences, requiring no LLM.
50%UnverifiedThe author believes that closed-source API solutions are easier to demo but harder to earn trust with, and for tools that touch real data, prefers debuggable open-source models.
50%