TrainCLI

Train ML models from a single command — no code required.

TrainCLI automates the entire machine learning pipeline: data profiling, cleaning, model selection, training, and export. Go from a CSV to a production-ready model in one command.

train --data sales.csv --target revenue

---

Installation

pip install traincli

Requirements: Python 3.9+, pandas, scikit-learn

Extra dependencies (e.g. xgboost) are installed on demand when you first use a model that needs them.

---

Quick Start

1. Train a model

train --data sales.csv --target revenue

TrainCLI will profile your data, clean it automatically, detect whether it's a classification or regression problem, run cross-validation across multiple algorithms, pick the best one, and export everything to ./output/.

2. Use your model

import pickle, pandas as pd

with open("output/model.pkl", "rb") as f:
    model = pickle.load(f)

prediction = model.predict(pd.DataFrame([{"feature1": 10, "feature2": "A"}]))

Or re-run the generated train.py to retrain from scratch at any time.

---

Features

One-command training No boilerplate. Point it at a CSV and a target column.

Automatic model selection Cross-validation across a full candidate pool — classification and regression. A bundled meta-model narrows the search based on dataset characteristics before confirming via CV.

Smart data cleaning Auto-handles ID columns, date parsing, currency strings, boolean encoding, and high-null columns. Prompts you only when a decision is genuinely ambiguous.

Leakage detection Scans for features with >0.98 correlation to the target, name-based hints, and linear transformations. Flags them before training, not after.

Full reproducibility Every training run exports a standalone train.py and a session.json audit trail. Re-run, share, or customize without touching TrainCLI again.

---

Architecture

CSV Input
   │
   ▼
Profiler ─────────────────────────────┐
   │  (type detection, leakage scan)   │
   ▼                                   │
Cleaner                               LLM Insights
   │  (auto-fix + interactive)         │  (optional)
   ▼                                   │
Detector ◄────────────────────────────┘
   │  (classification vs regression)
   ▼
Trainer
   │  (cross-validation, meta-model, selection)
   ▼
Exporter
   │
   ├── model.pkl
   ├── train.py
   ├── report.md
   ├── metrics.json
   └── session.json

Module responsibilities:

Module	Role
profiler.py	Column type detection, data quality issues, leakage analysis
cleaner.py	Auto-fixes and interactive prompts for ambiguous cases
detector.py	Heuristic problem-type detection with LLM fallback
trainer.py	Model selection, cross-validation, meta-model integration
exporter.py	Output generation — pkl, scripts, reports, metadata
mcp_server.py	MCP server exposing TrainCLI to Claude Desktop
core.py	Shared pipeline logic used by both CLI and MCP

---

Commands

Command	Description
train --data <file> --target <col>	Train a model on a CSV
train --dir <folder> --target <col>	Train on all CSVs in a folder
train --model <name>	Use a specific algorithm (e.g. random_forest)
train --model ?	Pick a model interactively from a ranked table
train --task classification	Override auto-detected problem type
train --drop col1,col2	Drop columns before training (leakage fix)
train --preview	Profile data without training
train --auto	Skip all prompts — useful in CI/CD

Examples:

# Force a specific model
train --data sales.csv --target price --model gradient_boosting

# Interactive model picker
train --data data.csv --target label --model ?

# Drop leaky columns
train --data jobs.csv --target salary --drop min_salary,max_salary

# Batch processing, no prompts
train --dir ./datasets/ --target outcome --auto

# Preview only
train --data data.csv --target col --preview

---

Supported Models

TrainCLI keeps the base install light — only pandas and scikit-learn are required. XGBoost is supported but not bundled; if cross-validation picks it as the best model for your data, TrainCLI will prompt you to install it before proceeding.

Classification: Logistic Regression, Random Forest, Gradient Boosting, SVM, K-Nearest Neighbors, XGBoost*

Regression: Linear Regression, Ridge Regression, Random Forest, Gradient Boosting, Support Vector Regression, XGBoost*

* Requires pip install xgboost. TrainCLI will prompt you if it's needed.

---

Output Files

File	Contents
model.pkl	Trained scikit-learn pipeline, ready to use
train.py	Standalone retraining script with educational comments
report.md	Human-readable summary — metrics, interpretation, decisions
metrics.json	Machine-readable metrics for downstream tooling
session.json	Full audit trail: every decision the pipeline made

---

MCP Server (Claude Desktop Integration)

TrainCLI ships with an MCP server so Claude Desktop can train models directly.

Setup:

pip install traincli

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "traincli": {
      "command": "traincli-mcp"
    }
  }
}

Restart Claude Desktop.

Available tools:

Tool	Description
train_model	Run the full pipeline on a CSV
profile_dataset	Analyze data without training
suggest_target	Recommend the best target column

Example:

User: Train a model on sales.csv to predict revenue

Claude: [calls train_model]
✓ Model trained: Gradient Boosting
✓ R² = 0.847
✓ Files saved to ./output/

---

Development

git clone https://github.com/yourusername/traincli
cd traincli
pip install -e .

---

License

MIT

Contributing

Open an issue or PR — contributions welcome.

---

Built for the IBM Dev Day: Bob Edition Hackathon.