mostlyai-engine 1.5.1

Synthetic Data Engine

Synthetic Data Engine 💎

Documentation | Technical Paper | Free Cloud Service

Create high-fidelity privacy-safe synthetic data:

1. prepare, analyze, and encode original data
2. train a generative model on the encoded data
3. generate synthetic data samples to your needs:
   • up-sample / down-sample
   • conditionally generate
   • rebalance categories
   • impute missings
   • incorporate fairness
   • adjust sampling temperature

...all within your safe compute environment, all with a few lines of Python code 💥.

Note: This library is the underlying model engine of the Synthetic Data SDK. Please refer to the latter, for an easy-to-use, higher-level software toolkit.

Installation

The latest release of mostlyai-engine can be installed via pip:

pip install -U mostlyai-engine

or alternatively for a GPU setup (needed for LLM finetuning and inference):

pip install -U 'mostlyai-engine[gpu]'

On Linux, one can explicitly install the CPU-only variant of torch together with mostlyai-engine:

pip install -U torch==2.7.0+cpu torchvision==0.22.0+cpu mostlyai-engine --extra-index-url https://download.pytorch.org/whl/cpu

Quick start

Tabular Model: flat data, without context

from pathlib import Path
import pandas as pd
from mostlyai import engine

# set up workspace and default logging
ws = Path("ws-tabular-flat")
engine.init_logging()

# load original data
url = "https://github.com/mostly-ai/public-demo-data/raw/refs/heads/dev/census"
trn_df = pd.read_csv(f"{url}/census.csv.gz")

# execute the engine steps
engine.split(                         # split data as PQT files for `trn` + `val` to `{ws}/OriginalData/tgt-data`
  workspace_dir=ws,
  tgt_data=trn_df,
  model_type="TABULAR",
)
engine.analyze(workspace_dir=ws)      # generate column-level statistics to `{ws}/ModelData/tgt-stats/stats.json`
engine.encode(workspace_dir=ws)       # encode training data to `{ws}/OriginalData/encoded-data`
engine.train(                         # train model and store to `{ws}/ModelStore/model-data`
    workspace_dir=ws,
    max_training_time=1,              # limit TRAIN to 1 minute for demo purposes
)
engine.generate(workspace_dir=ws)     # use model to generate synthetic samples to `{ws}/SyntheticData`
pd.read_parquet(ws / "SyntheticData") # load synthetic data

Tabular Model: sequential data, with context

from pathlib import Path
import pandas as pd
from mostlyai import engine

engine.init_logging()

# set up workspace and default logging
ws = Path("ws-tabular-sequential")
engine.init_logging()

# load original data
url = "https://github.com/mostly-ai/public-demo-data/raw/refs/heads/dev/baseball"
trn_ctx_df = pd.read_csv(f"{url}/players.csv.gz")  # context data
trn_tgt_df = pd.read_csv(f"{url}/batting.csv.gz")  # target data

# execute the engine steps
engine.split(                         # split data as PQT files for `trn` + `val` to `{ws}/OriginalData/(tgt|ctx)-data`
  workspace_dir=ws,
  tgt_data=trn_tgt_df,
  ctx_data=trn_ctx_df,
  tgt_context_key="players_id",
  ctx_primary_key="id",
  model_type="TABULAR",
)
engine.analyze(workspace_dir=ws)      # generate column-level statistics to `{ws}/ModelStore/(tgt|ctx)-data/stats.json`
engine.encode(workspace_dir=ws)       # encode training data to `{ws}/OriginalData/encoded-data`
engine.train(                         # train model and store to `{ws}/ModelStore/model-data`
    workspace_dir=ws,
    max_training_time=1,              # limit TRAIN to 1 minute for demo purposes
)
engine.generate(workspace_dir=ws)     # use model to generate synthetic samples to `{ws}/SyntheticData`
pd.read_parquet(ws / "SyntheticData") # load synthetic data

Language Model: flat data, without context

from pathlib import Path
import pandas as pd
from mostlyai import engine

# init workspace and logging
ws = Path("ws-language-flat")
engine.init_logging()

# load original data
trn_df = pd.read_parquet("https://github.com/mostly-ai/public-demo-data/raw/refs/heads/dev/headlines/headlines.parquet")
trn_df = trn_df.sample(n=10_000, random_state=42)

# execute the engine steps
engine.split(                         # split data as PQT files for `trn` + `val` to `{ws}/OriginalData/tgt-data`
    workspace_dir=ws,
    tgt_data=trn_df,
    tgt_encoding_types={
        'category': 'LANGUAGE_CATEGORICAL',
        'date': 'LANGUAGE_DATETIME',
        'headline': 'LANGUAGE_TEXT',
    }
)
engine.analyze(workspace_dir=ws)      # generate column-level statistics to `{ws}/ModelStore/tgt-stats/stats.json`
engine.encode(workspace_dir=ws)       # encode training data to `{ws}/OriginalData/encoded-data`
engine.train(                         # train model and store to `{ws}/ModelStore/model-data`
    workspace_dir=ws,
    max_training_time=2,                   # limit TRAIN to 2 minute for demo purposes
    model="MOSTLY_AI/LSTMFromScratch-3m",  # use a light-weight LSTM model, trained from scratch (GPU recommended)
    # model="microsoft/phi-1.5",           # alternatively use a pre-trained HF-hosted LLM model (GPU required)
)
engine.generate(                      # use model to generate synthetic samples to `{ws}/SyntheticData`
    workspace_dir=ws,
    sample_size=10,
)
pd.read_parquet(ws / "SyntheticData") # load synthetic data