Data Preparation

prepare_for_training is the main entry point for preparing Lightning Rod datasets for model training. It filters invalid samples, deduplicates, splits into train/test, and returns SampleDataset objects ready for lr.training.run(...) with a GRPOTrainingConfig or SFTTrainingConfig, for lr.evals.run_from_training_job(config, job, test_dataset) (GRPO), or for lr.evals.create(...) when you supply the full model list (e.g. SFT or custom evals).

What It Does

1. Filter — Drops invalid samples, optionally by resolution horizon and context presence

2. Deduplicate — Removes duplicates by (question_text, resolution_date) or custom key

3. Split — Splits into train/test by temporal order or random shuffle

Parameters

from lightningrod.training import prepare_for_training, FilterParams, DedupParams, SplitParams

train_dataset, test_dataset = prepare_for_training(
    dataset,
    filter=FilterParams(...),
    dedup=DedupParams(...),
    split=SplitParams(...),
    verbose=True,
)

Parameter	Type	Default	Description
dataset	SampleDataset	—	Dataset to prepare (required)
filter	FilterParams | None	None	Filtering options (see below)
dedup	DedupParams | None	None	Deduplication options (see below)
split	SplitParams | None	None	Train/test split options (see below)
verbose	bool	True	Print filter/dedup/split stats

FilterParams

Controls which samples are kept before splitting.

Field	Type	Default	Description
days_to_resolution_range	(min, max) | None	None	Filter by resolution horizon. E.g. (90, None) keeps ≥ 90 days; (14, 60) keeps 14–60 days
drop_missing_context	bool	False	Exclude samples with no rendered context

DedupParams

Controls duplicate removal.

Field	Type	Default	Description
key_fn	Callable | None	None	Custom dedup key; default is (question_text, resolution_date)

SplitParams

Controls how samples are divided into train and test sets.

Field	Type	Default	Description
strategy	str	"temporal"	"temporal" (chronological) or "random"
test_size	float | None	0.2	Fraction of samples for test set (0.0–1.0)
test_start	str | None	None	ISO date for temporal split cutoff; alternative to test_size for exact date control
random_state	int	196	Seed for reproducible random splits
filter_leaky_train	bool	True	Remove train samples whose date_close or resolution_date falls within the test period
sort_key	Callable | None	None	Custom sort key for temporal ordering; defaults to resolution_date
leakage_keys	list[Callable] | None	None	Additional fields to check for temporal leakage

Guidelines

• Prefer "temporal" split for forecasting — It reflects real-world deployment where the model never sees future questions during training.

• Use test_start vs test_size — Use test_start when you want the test set to cover a specific time window (e.g. a recent month); use test_size when you just want a fixed proportion.

• Use days_to_resolution_range to control difficulty — Short horizons (e.g. (7, 30)) are easier; longer horizons (e.g. (90, None)) test genuine uncertainty.

• Set drop_missing_context=True only if your dataset was generated with context_generators (e.g. NewsContextGenerator) in QuestionPipeline. Context is absent by default, so enabling this on a dataset without context will drop all samples. See Labeling and Context.

• Keep filter_leaky_train=True (default) unless you intentionally want to test without temporal leak protection.

• Use verbose=True on first runs to understand how many samples are dropped at each stage and why.

Example

from lightningrod.training import prepare_for_training, FilterParams, SplitParams

train_dataset, test_dataset = prepare_for_training(
    dataset,
    filter=FilterParams(days_to_resolution_range=(90, None)),
    split=SplitParams(test_size=0.2),
    verbose=True,
)

Output

Returns (train_dataset, test_dataset) — two SampleDataset objects ready for lr.training.run(training_config, dataset=train_dataset) (GRPO or SFT config) and evaluation on test_dataset via lr.evals.run_from_training_job(training_config, job, test_dataset) or lr.evals.create(...).