smolmodels 0.9.2

A framework for building ML models from natural language

smolmodels ✨

Build machine learning models using natural language and minimal code

Quickstart | Features | Installation & Setup | Documentation | Benchmarks

Create machine learning models with minimal code by describing what you want them to do in plain words. You explain the task, and the library builds a model for you, including data generation, feature engineering, training, and packaging.

[!NOTE] This library is in early development, and we're actively working on new features and improvements! Please report any bugs or share your feature requests on GitHub or Discord 💛

1. Quickstart

Installation:

pip install smolmodels

Define, train and save a Model:

import smolmodels as sm

# Step 1: define the model
model = sm.Model(
    intent="Predict sentiment on a news article such that [...]",
    input_schema={"headline": str, "content": str},         # [optional - can be pydantic or dict]
    output_schema={"sentiment": str}                        # [optional - can be pydantic or dict]
)

# Step 2: build and train the model on data
model.build(
   datasets=[dataset, auxiliary_dataset],
   provider="openai/gpt-4o-mini",
   timeout=3600
)

# Step 3: use the model to get predictions on new data
sentiment = model.predict({
   "headline": "600B wiped off NVIDIA market cap",
   "content": "NVIDIA shares fell 38% after [...]",
})

# Step 4: save the model, can be loaded later for reuse
sm.save_model(model, "news-sentiment-predictor")

# Step 5: load a saved model and use it
loaded_model = sm.load_model("news-sentiment-predictor.tar.gz")

2. Features

smolmodels combines graph search, LLM code/data generation and code execution to produce a machine learning model that meets the criteria of the task description. When you call model.build(), the library generates a graph of possible model solutions, evaluates them, and selects the one that maximises the performance metric for this task.

2.1. 💬 Define Models using Natural Language

A model is defined as a transformation from an input schema to an output schema, which behaves according to an intent. The schemas can be defined either using pydantic models, or plain dictionaries that are convertible to pydantic models.

# This defines the model's identity
model = sm.Model(
    intent="Predict sentiment on a news article such that [...]",
    input_schema={"headline": str, "content": str},                 # supported: pydantic or dict
    output_schema={"sentiment": str}                                # supported: pydantic or dict
)

You describe the model's expected behaviour in plain English. The library will select a metric to optimise for, and produce logic for feature engineering, model training, evaluation, and so on.

2.2. 🎯 Model Building

The model is built by calling model.build(). This method takes one or more datasets and generates a set of possible model solutions, training and evaluating them to select the best one. The model with the highest performance metric becomes the "implementation" of the predictor.

You can specify the model building cutoff in terms of a timeout, a maximum number of solutions to explore, or both.

model.build(
    datasets=[dataset_a, dataset_b],
    provider="openai/gpt-4o-mini",
    timeout=3600,                       # [optional] max time in seconds
    max_iterations=10                   # [optional] max number of model solutions to explore
)

The model can now be used to make predictions, and can be saved or loaded using sm.save_model() or sm.load_model().

sentiment = model.predict({"headline": "600B wiped off NVIDIA market cap", ...})

2.3. 🎲 Data Generation and Schema Inference

The library can generate synthetic data for training and testing. This is useful if you have no data available, or want to augment existing data. You can do this with the sm.DatasetGenerator class:

dataset = sm.DatasetGenerator(
    schema={"headline": str, "content": str, "sentiment": str},  # supported: pydantic or dict
    data=existing_data
)
dataset.generate(1000)

model.build(
    datasets=[dataset],
    ...
)

[!CAUTION] Data generation can consume a lot of tokens. Start with a conservative generate_samples value and increase it if needed.

The library can also infer the input and/or output schema of your predictor, if required. This is based either on the dataset you provide, or on the model's intent. This can be useful when you don't know what the model should look like. As with the models, you can specify the schema using pydantic models or plain dictionaries.

# In this case, the library will infer a schema from the intent and generate data for you
model = sm.Model(intent="Predict sentiment on a news article such that [...]")
model.build(provider="openai/gpt-4o-mini")

[!TIP] If you know how the model will be used, you will get better results by specifying the schema explicitly. Schema inference is primarily intended to be used if you don't know what the input/output schema at prediction time should be.

2.4. 🌐 Multi-Provider Support

You can use multiple LLM providers for model generation. Specify the provider and model in the format provider/model:

model.build(provider="openai/gpt-4o-mini", ...)

See the section on installation and setup for more details on supported providers and how to configure API keys.

3. Installation & Setup

Install the library in the usual manner:

pip install smolmodels

Set your API key as an environment variable based on which provider you want to use. For example:

# For OpenAI
export OPENAI_API_KEY=<your-API-key>
# For Anthropic
export ANTHROPIC_API_KEY=<your-API-key>
# For Gemini
export GEMINI_API_KEY=<your-API-key>

[!TIP] The library uses LiteLLM as its provider abstraction layer. For other supported providers and models, check the LiteLLM documentation.

4. Documentation

For full documentation, visit docs.plexe.ai.

5. Benchmarks

Performance evaluated on 20 OpenML benchmark datasets and 12 Kaggle competitions. Higher performance observed on 12/20 OpenML datasets, with remaining datasets showing performance within 0.005 of baseline. Experiments conducted on standard infrastructure (8 vCPUs, 30GB RAM) with 1-hour runtime limit per dataset.

Complete code and results are available at plexe-ai/plexe-results.

6. Contributing

We love contributions! You can get started with issues, submitting a PR with improvements, or joining the Discord to chat with the team. See CONTRIBUTING.md for detailed guidelines.

7. License

Apache-2.0 License - see LICENSE for details.

8. Product Roadmap

• Fine-tuning and transfer learning for small pre-trained models
• Support for non-tabular data types in model generation
• Use Pydantic for schemas and split data generation into a separate module
• Smolmodels self-hosted platform ⭐ (More details coming soon!)