TextPredictionConfig

Bases: TrainerConfig

Configuration for TextPredictionTrainer.

Unified data contract: internal representation uses masked, split, factual_class, alternative_class, factual, alternative, transition. Training uses only rows where transition ∈ {c1→c2, c2→c1} for the configured pair.

Data input:

• data as Dict[str, DataFrame]: per-class format. Key = factual_class; each df has

masked, split, and columns = class names (factual = df[factual_class], alternative = df[alternative_class]). If a df has no column for another class (single-token-per-class, e.g. Gender), target is inferred as the other class's token for the configured pair.

• data as DataFrame: merged format with label_class_col, label_col, and optionally

target_col, target_class_col. If target columns omitted, pair is required and target = other class's token.

• hf_dataset: load from HuggingFace (merged format), then convert to unified.

Attributes:

Name	Type	Description
run_id	Optional[str]	Optional run identifier (subdir and display).
data	Optional[Union[DataFrame, Dict[str, DataFrame], str, Path]]	Per-class dict (label_class -> DataFrame) or merged DataFrame.
hf_dataset	Optional[str]	HuggingFace dataset ID; loads merged format and converts to unified.
hf_subset	Optional[Union[str, List[str]]]	Subset name(s) to load when using hf_dataset.
hf_splits	Optional[List[str]]	Splits to include (e.g. ['train', 'validation', 'test']).
target_classes	Optional[List[str]]	Target classes for training. Pair is automatically inferred when len(target_classes) == 2.
all_classes	Optional[List[str]]	All classes available in the dataset. If None (default), inferred from data. When loading from HuggingFace datasets, None means load all configs/subsets.
masked_col,	split_col	Column names. For merged format also label_col, label_class_col, and optionally target_col, target_class_col.
use_class_names_as_columns	bool	For per-class data, use class name as column name for tokens.

all_classes class-attribute instance-attribute

all_classes = None

All classes available in the dataset. If None (default), inferred from data. When loading from HuggingFace datasets, None means load all configs/subsets.

alternative_class_col class-attribute instance-attribute

alternative_class_col = 'alternative_class'

alternative_col class-attribute instance-attribute

alternative_col = 'alternative'

class_merge_map class-attribute instance-attribute

class_merge_map = None

class_merge_transition_groups class-attribute instance-attribute

class_merge_transition_groups = None

data class-attribute instance-attribute

data = None

dataset_trust_remote_code class-attribute instance-attribute

dataset_trust_remote_code = None

Optional trust_remote_code value for HuggingFace datasets.load_dataset. None means do not pass the keyword. This is separate from TrainingArguments.trust_remote_code, which applies to models/tokenizers.

decoder_eval_export_row_wise_csv class-attribute instance-attribute

decoder_eval_export_row_wise_csv = False

decoder_eval_ignore_tokens class-attribute instance-attribute

decoder_eval_ignore_tokens = None

decoder_eval_lms_max_samples class-attribute instance-attribute

decoder_eval_lms_max_samples = None

decoder_eval_prob_on_other_class class-attribute instance-attribute

decoder_eval_prob_on_other_class = True

decoder_eval_restrict_to_target_classes class-attribute instance-attribute

decoder_eval_restrict_to_target_classes = True

decoder_eval_targets class-attribute instance-attribute

decoder_eval_targets = None

eval_neutral_additional_excluded_words class-attribute instance-attribute

eval_neutral_additional_excluded_words = None

eval_neutral_data class-attribute instance-attribute

eval_neutral_data = None

eval_neutral_max_rows class-attribute instance-attribute

eval_neutral_max_rows = None

hf_dataset class-attribute instance-attribute

hf_dataset = None

hf_splits class-attribute instance-attribute

hf_splits = None

hf_subset class-attribute instance-attribute

hf_subset = None

label_class_col class-attribute instance-attribute

label_class_col = 'label_class'

label_col class-attribute instance-attribute

label_col = 'label'

masked_col class-attribute instance-attribute

masked_col = 'masked'

max_counterfactuals_per_sentence class-attribute instance-attribute

max_counterfactuals_per_sentence = 1

n_features class-attribute instance-attribute

n_features = 1

random_state class-attribute instance-attribute

random_state = None

run_id class-attribute instance-attribute

run_id = None

split_col class-attribute instance-attribute

split_col = 'split'

Dataset split column. "split" (default) uses row-level splits from the data. None assigns random row-level train/validation/test splits (reshuffled when split_resplit_per_seed=True). "heldout" assigns vocabulary-held-out splits by factual token (per target class). Vocabulary-held-out splitting is never selected automatically; the user must explicitly set split_col="heldout".

split_group_col class-attribute instance-attribute

split_group_col = None

Unified column for vocabulary-held-out resplit (default: factual token column).

split_group_key class-attribute instance-attribute

split_group_key = None

Callable or sequence applied to split_group_col before grouping (e.g. [str.strip, str.casefold]).

split_ratios class-attribute instance-attribute

split_ratios = None

Train/validation/test fractions as (train, val, test) or {train, validation, test}.

split_test_ratio class-attribute instance-attribute

split_test_ratio = 0.1

split_train_ratio class-attribute instance-attribute

split_train_ratio = 0.8

split_val_ratio class-attribute instance-attribute

split_val_ratio = 0.1

target_classes class-attribute instance-attribute

target_classes = None

Target classes for training. Pair is automatically inferred when len(target_classes) == 2.

use_class_names_as_columns class-attribute instance-attribute

use_class_names_as_columns = True

__post_init__

__post_init__()

Normalize split-ratio inputs after dataclass initialization.

Source code in gradiend/trainer/text/prediction/trainer.py

def __post_init__(self) -> None:
    """Normalize split-ratio inputs after dataclass initialization."""
    train, val, test = normalize_split_ratios(
        self.split_ratios,
        train=self.split_train_ratio,
        val=self.split_val_ratio,
        test=self.split_test_ratio,
    )
    object.__setattr__(self, "split_train_ratio", train)
    object.__setattr__(self, "split_val_ratio", val)
    object.__setattr__(self, "split_test_ratio", test)

__str__

__str__()

Source code in gradiend/trainer/text/prediction/trainer.py

def __str__(self) -> str:
    return (
        f"TextPredictionConfig(img_format={self.img_format!r}, target_classes={self.target_classes!r}, "
        f"masked_col={self.masked_col!r}, n_features={self.n_features})"
    )