ForgeBaselines

Dockerized ML orchestration platform for generating reproducible baselines on tabular classification and information retrieval tasks. Upload a CSV, configure your experiment, get a ranked leaderboard.

Built with FastAPI · scikit-learn · BM25 · MLflow · Next.js · Docker

---

Architecture

  User ───────────▶ Frontend (Next.js) — localhost:3000
                         │ API calls
                         ▼
                    Orchestrator (FastAPI, 8000)
                    ├──▶ Classification (FastAPI, 8001)
                    ├──▶ IR / BM25 (FastAPI, 8002)
                    └──▶ MLflow (5001)

All services run locally via Docker Compose. There is no cloud deployment at this time.

Service	Port	Role
frontend	3000	Next.js UI — upload, configure, progress, results
orchestrator	8000	Main API — routing, preprocessing, experiment orchestration
classification	8001	Internal — model training + MLflow logging
ir	8002	Internal — BM25 retrieval + metrics + MLflow logging
mlflow	5001	Experiment tracking UI + REST API

Data is stored locally under ./data/ (Docker volume mount). MLflow artifacts are stored locally under ./mlflow-artifacts/.

---

Prerequisites

• Docker Desktop

---

Local Setup

git clone https://github.com/Parth-Agarwal216/ForgeBaselines.git
cd ForgeBaselines
cp .env.example .env
docker compose up --build   # first build ~3-5 min

URL	What
http://localhost:3000	Main UI
http://localhost:8000/docs	Orchestrator API (Swagger)
http://localhost:5001	MLflow experiment tracker

---

Usage

Classification

1. Upload — drop a CSV at /upload
2. Configure — pick target column, review auto-detected column roles, select models, set preprocessing options
3. Run — trains selected models (Logistic Regression, Random Forest, Gradient Boosting, XGBoost, SVM, KNN)
4. Results — live progress bar while training, leaderboard ranked by F1 on completion, download as CSV

Information Retrieval

1. Upload — upload a corpus CSV and a queries CSV from the dashboard
2. Configure — select corpus/query datasets and map text columns
3. Run — builds a BM25 index and scores all queries
4. Results — live progress bar during scoring, MAP / nDCG@10 / Recall / MRR metrics on completion

Both experiment types run asynchronously — the POST returns immediately and the results page polls /status until completion.

---

Seed Data

pip install pandas scikit-learn requests   # one-time, host machine
python scripts/seed_test_data.py

Uploads three sample datasets (Iris, Wine, Titanic) so you can test the full pipeline immediately.

---

Tests

docker compose exec orchestrator pytest tests/ -v --tb=short
docker compose exec classification pytest tests/ -v --tb=short

---

CI

Every push to main triggers GitHub Actions CI: pytest for both backend services + npm run build for the frontend (3 parallel jobs).

---

Development

Hot reload is on for orchestrator and frontend. The IR service requires a rebuild after code changes:

docker compose up -d --build ir

Other useful commands:

docker compose logs -f orchestrator    # tail a service
docker compose up --build frontend     # rebuild after dep changes
docker compose down                    # stop everything

---

Auth

Sign in with a magic link (passwordless). Enter your email → click the link → you're in. Each user's datasets and experiments are fully isolated.