Repository navigation
Mock data generation — no fit() required. New model.mock(schema=...) builds plausible rows from a declarative schema instead of real training data. Per-column specs support integer, precision, dist="normal" with mean/std, categorical weights, and null_prob for realistic missing values. The existing conditions= operators work unchanged — schema declarations are implicit constraints, conditions= tighten them. Reproducible via seed=. Few-shot grounding via examples= (rows prepended as in-context demonstrations — big realism boost on LLMs that don't natively know the GReaT row format). Pairs well with tabularisai/Qwen3-0.3B-distil. Save/share schemas via save_schema / load_schema (JSON or YAML).
from be_great import GReaT
from be_great.great_mock_datasets import save_schema, load_schema
schema = {
"age": {"type": "num", "range": (18, 99), "integer": True,
"dist": "normal", "mean": 38, "std": 13, "null_prob": 0.05},
"sex": {"type": "cat", "values": ["Male", "Female", "Other"],
"weights": [0.5, 0.45, 0.05]},
"income": {"type": "cat", "values": ["<=50K", ">50K"], "weights": [0.76, 0.24]},
}
model = GReaT(llm="tabularisai/Qwen3-0.3B-distil")
mock = model.mock(schema=schema, n_samples=100, seed=42,
conditions={"age": ">= 40", "sex": "!= 'Male'"})chugchug as the default progress bar for both training and sampling, replacing HF's tqdm. The new ChugProgressCallback shows a gradient bar with auto-colored metrics — loss goes green when improving, red when worsening — plus live lr and epoch. Notebooks get HTML/CSS gradient bars automatically. chugchug >= 0.1.3 is now a required dependency.
Live synthesis-quality monitoring during training. Pass eval_data= to fit() and, after each epoch (or every N steps via eval_every=), GReaT samples from the in-training model, scores it with ColumnShapes, and pushes q=… into the chugchug bar. A 🎯 [quality @ step N] column_shapes_mean = 0.XXXX line is also printed. Watch synthesis quality climb in real time — training signal beyond raw loss.
Auto device detection. Generation methods now default to device="auto": CUDA → MPS → CPU. Apple Silicon works out of the box. dataloader_pin_memory defaults to False on non-CUDA backends, silencing the MPS warning.
First-token bias on constrained sampling. When the schema declares dist="normal" or weights, the trie's first valid token is biased by the value-weight distribution — pulling generations toward the declared shape instead of uniform over the constraint set.
Fix: garbage tails in guided sampling output ("Female, age is 73"). Parser now picks the earliest delimiter among ;, ,, \n (plus . for non-numeric columns) instead of ; first. Numeric columns preserve decimals.
minor: New be_great/great_mock_datasets.py cleanly separates the mock path from the fit path. Added __repr__ to GReaT. Per-column col_stats["precision"] override in _enumerate_numeric. Version bumped 0.0.12 → 0.0.14 so source matches the PyPI release (0.0.13 was tagged but never source-bumped).