Apppublic

weiqi001/bankruptcy-mlflow-dashboard

sourceHugging Facemitupdated 2mo agoView on Hugging Face

0likes

App README

📊 MLflow tracking dashboard

The experiment history for the Corporate Bankruptcy Early-Warning System, served from Hugging Face rather than from someone's laptop.

Every training run of the pipeline records ~15 parameters, ~15 metrics and its artefacts, plus a nested run for each candidate algorithm on each feature set. That is what you are looking at.

Where the data comes from

GitHub Actions                          Hugging Face
--------------                          ------------
model-training
  mlflow server on :5555 (in the job)
  train.py logs every run
        │
        ▼
sync-tracking
  uploads mlflow.db + mlruns/  ────────▶ dataset repo
  restarts this Space          ────────▶ this Space
                                            │
                                            ▼
                                    restore.py pulls the
                                    snapshot on boot, then
                                    mlflow server serves it

The dashboard therefore reflects every completed pipeline run, refreshed within about a minute of one finishing, because the pipeline restarts this Space after it syncs.

What "real-time" means here

Mode	What you see	Cost
Snapshot (default)	Every run, refreshed after each pipeline run	Nothing; the Space stays read-only in practice
Live (optional)	Metrics appearing during a training run	The pipeline must log straight to this Space; see below

To switch to live logging, point the training job at this Space:

yaml

env:
  MLFLOW_TRACKING_URI: https://<owner>-<space>.hf.space

Read the trade-offs in the project's docs/MLFLOW_DASHBOARD.md before doing that — chiefly that a Space filesystem is ephemeral, so a restart loses anything not yet synced, and that a public Space is writable by anyone who knows the URL.

Configuration

Set in the Space's Settings → Variables and secrets:

Name	Kind	Value
`MLFLOW_DATASET_REPO`	variable	`<owner>/bankruptcy-early-warning-mlflow`
`HF_TOKEN`	secret	only if that dataset repo is private

Without MLFLOW_DATASET_REPO the Space still starts, with an empty store.

—Application — the Streamlit early-warning app
—Dataset — the 10,503 firm-years the model was fitted on
—Source — the GitHub repository and its MLOps pipeline

📊 MLflow tracking dashboard

Where the data comes from

What "real-time" means here

Configuration

Related