modelgpt 0.2.2

LLM-powered AutoML — model selection, tuning, and training via any LLM

ModelGPT

LLM-powered AutoML. Describe your dataset, pick a task — ModelGPT asks an LLM to design, tune, and fit the best model for you. Works with OpenAI, Anthropic, Groq, Ollama, and any other LiteLLM-supported provider.

from modelgpt import ModelGPT

mg = ModelGPT(model="gpt-4o-mini", api_key="sk-...")
model = mg.fit(X_train, y_train, task="regression")
predictions = model.predict(X_test)

---

How it works

Your data  ──▶  Dataset summary  ──▶  LLM prompt
                                          │
                                          ▼
                                 LLM generates Python
                                 (ensemble + CV tuning)
                                          │
                                          ▼
                                   Sandboxed execution
                            (subprocess isolation, timeout=120s,
                            crash & resource containment)
                                          │
                             ┌────────────┴────────────┐
                             ▼                         ▼
                       Fitted model           Error? retry (→ LLM
                       returned               sees error, self-corrects)
                                                       │
                                              Max retries hit?
                                                       ▼
                                              Safe fallback model

ModelGPT feeds the LLM a structured dataset summary (shape, dtypes, missing values, target statistics, class distribution). The LLM returns raw Python that uses cross-validated hyperparameter search and an ensemble method. That code runs in a controlled namespace with X and y already bound. If it fails, the error is sent back to the LLM for self-correction — up to max_retries times. If every retry fails, ModelGPT falls back to a safe, well-tuned GradientBoosting model so .fit() never crashes your pipeline.

---

Installation

Install the base package plus whichever backend you want to use:

pip install modelgpt[openai]      # OpenAI (gpt-4o, gpt-4o-mini, …)
pip install modelgpt[anthropic]   # Anthropic (claude-haiku-4-5, claude-opus-4-5, …)
pip install modelgpt[groq]        # Groq (fast + free tier)
pip install modelgpt[ollama]      # Ollama (local, no API key needed)
pip install modelgpt[all]         # everything including xgboost, lightgbm, catboost

---

Quickstart

OpenAI

from modelgpt import ModelGPT

mg = ModelGPT(model="gpt-4o-mini", api_key="sk-...")
model = mg.fit(X_train, y_train, task="regression", metric="RMSE")
predictions = model.predict(X_test)

Anthropic

mg = ModelGPT(model="claude-haiku-4-5", api_key="sk-ant-...")
model = mg.fit(X_train, y_train, task="classification", metric="AUC")

Groq

mg = ModelGPT(model="groq/llama3-70b-8192", api_key="gsk_...")
model = mg.fit(X_train, y_train, task="classification")

Ollama (local, no API key)

# First: install Ollama (https://ollama.com/download), then:
ollama pull qwen2.5-coder:7b
ollama serve

mg = ModelGPT(model="ollama/qwen2.5-coder:7b")
model = mg.fit(X_train, y_train, task="regression")

Using environment variables instead of passing api_key

export OPENAI_API_KEY="sk-..."
export ANTHROPIC_API_KEY="sk-ant-..."
export GROQ_API_KEY="gsk_..."

mg = ModelGPT(model="gpt-4o-mini")   # picks up key from env automatically

---

API Reference

ModelGPT(model, api_key, max_retries, verbose)

Parameter	Type	Default	Description
model	str	"gpt-4o-mini"	Any LiteLLM model string (see examples above)
api_key	str | None	None	API key for the backend. If None, reads from environment variable
max_retries	int	3	How many times to retry + self-correct on code execution failure
verbose	bool	True	Print generated code and progress to stdout

.fit(X, y, task, metric)

Parameter	Type	Default	Description
X	pd.DataFrame	—	Feature matrix
y	pd.Series	—	Target vector
task	str	"regression"	"regression" or "classification"
metric	str | None	None	Metric hint passed to the LLM (e.g. "RMSE", "AUC", "F1")

Returns a fitted sklearn-compatible model with a .predict() method.

---

Choosing a model

Larger and more capable models tend to generate better, more reliable code. Here are some good options ranked by capability:

Provider	Model string	Notes
OpenAI	"gpt-4o"	Best results
OpenAI	"gpt-4o-mini"	Fast, cheap, good default
Anthropic	"claude-opus-4-5"	Excellent code generation
Anthropic	"claude-haiku-4-5"	Fast and affordable
Groq	"groq/llama3-70b-8192"	Very fast, free tier available
Ollama	"ollama/qwen2.5-coder:7b"	Local, fully private
Ollama	"ollama/qwen2.5-coder:32b"	Local, best quality

---

Tips for better results

Tell the LLM which metric matters. The metric argument is forwarded directly into the prompt:

model = mg.fit(X_train, y_train, task="regression", metric="RMSE")
model = mg.fit(X_train, y_train, task="classification", metric="AUC")

Increase retries for hard tasks. If the LLM frequently generates broken code, raise max_retries:

mg = ModelGPT(model="gpt-4o-mini", max_retries=5)

Inspect what was generated. With verbose=True (the default) the full generated code is printed. You can copy it, tweak it, and re-run it manually.

Fallback is always safe. If every retry fails, ModelGPT automatically falls back to a well-tuned GradientBoostingRegressor / GradientBoostingClassifier, so .fit() never crashes your pipeline.

Install the boosting extras for the best LLM-generated models:

pip install modelgpt[boosting]   # adds xgboost, lightgbm, catboost

---

Project structure

modelGPT/
├── modelgpt/
│   ├── __init__.py          # Public API
│   └── modelgpt.py          # Core ModelGPT class
├── tests/                   # Test suite
├── example/
│   └── example_usage.py     # Quickstart demo (diabetes dataset)
├── pyproject.toml
└── README.md

---

Requirements

• Python ≥ 3.10
• litellm, pandas, scikit-learn, numpy (installed automatically)
• An API key for your chosen backend or Ollama running locally

---

License

MIT