naialorente/datalog

mltrackr is a lightweight Python library for tracking machine learning experiments. When you are training a model and want to record which settings produced the best results, mltrackr lets you do that with just two lines of code added to your existing training loop. It does not require a separate server, an online account, or any cloud service. Everything is saved in a single SQLite database file on your computer. After installing via pip, you wrap your training loop with a run context and call mltrackr.log() to record any numbers you care about. Once you have collected a few runs, you can view them in a local web dashboard at http://localhost:7000 that shows charts, metric comparisons, and trend indicators side by side. The dashboard auto-refreshes every five seconds, so you can watch your metrics update in real time while training. Beyond basic logging, mltrackr includes several tools aimed at saving time during experimentation. An anomaly detection feature watches for problems like the loss becoming not-a-number or a metric that has stopped improving for a set number of epochs. A hyperparameter suggestion command analyzes your run history and reports which settings correlated with better results in plain English. There is also a report generator that produces a markdown summary of all your experiments, and optionally uses a locally installed language model (Ollama) to write a narrative, with no cloud API keys required. The library works with most Python machine learning frameworks including PyTorch, scikit-learn, Keras, HuggingFace Transformers, XGBoost, and LightGBM. All data lives in a file at ~/.mltrackr/experiments.db that you can open with any SQLite browser or back up like any other file. The project is available on PyPI and is licensed under the MIT license. It targets developers and researchers who want quick experiment tracking without the overhead of tools like MLflow or Weights and Biases.