micm_nlp.config

CONFIG — the YAML schema, validated.

CONFIG.from_yaml loads a run description into typed sections: task, peft, model, tokenizer, ds, eval, test, trainer, training_args, data_collator, custom_training_args, cuda and env. Loading also applies the env block to os.environ.

Two design points shape everything here.

Sections accept extra keys. Every section inherits from _Flex, which allows extras, implements the mapping protocol so dict(obj) and **obj expose both declared fields and extras, and recursively wraps nested dicts. YAML may therefore carry keys the schema does not declare, and runtime code may attach computed attributes (uuid4, param_size). Note vars(obj) does not see extras — pydantic keeps them in __pydantic_extra__; use dict(obj).

Class selection stays in YAML. trainer.cls, training_args.cls, data_collator.cls and model.pretrained.cls are thin shells: the real schema lives in HuggingFace, and the matching args block is splatted into the constructor at runtime.

Importing this module also widens PyYAML’s float resolver, so 5e-5 parses as a float rather than a string — YAML 1.1 otherwise requires a decimal point.

Classes

AdapterConfig	Locates a saved PEFT adapter to load: by name or uuid4, optionally
CONFIG	One run, fully described.
CudaConfig	The cuda block. empty_cache_steps frees the allocator cache every N
CustomTrainingArgsConfig	Settings this package adds beyond HuggingFace's TrainingArguments.
DataCollatorConfig	Thin shell. The actual schema lives in HF (DataCollatorForLanguageModeling,
DatasetConfig	The ds block: which dataset, where it lives, and how to read it.
EvalConfig	The eval block: when evaluation runs.
InitConfig	TRAIN-from-scratch spec: which model class to instantiate and which
InitModelConfigConfig	HF Config class + its constructor kwargs. Used in TRAIN mode to build
InputConfig	Which dataset column(s) hold the input.
LabelConfig	Which column holds the label, and what the label space is.
ModelConfig	The model block: architecture, and how the model is obtained.
PeftConfig	The peft block: which PEFT method, and how it is parameterized.
PostprocConfig	task.preproc_rules: what happens to predictions before metrics see them.
PretrainedConfig	Locates a pretrained model, and names the class to load it with.
SplitsConfig	Which splits a dataset already ships with.
TaskConfig	The task block: what is being learned and how it is scored.
TaskIdConfig	Which column identifies the task, for multi-task runs that score each
TestConfig	The test block: whether and how the held-out test split is scored.
TokenizerConfig	The tokenizer block: which tokenizer to load, or which to train.
TrainerConfig	Thin shell. cls selects which HF Trainer subclass to instantiate
TrainingArgsConfig	Thin shell. The actual schema lives in HF (TrainingArguments,

Module Contents

class micm_nlp.config.AdapterConfig

Bases: _Flex

Locates a saved PEFT adapter to load: by name or uuid4, optionally a specific checkpoint, from source.

checkpoint: str | None = None

name: str | None = None

source: str | None = None

uuid4: str | None = None

class micm_nlp.config.CONFIG

Bases: _Flex

One run, fully described.

Every section is optional except mode, because the same schema covers training, evaluation, preprocessing and tokenizer training – a preprocessing config has no training_args, and validation only demands what the mode needs. See from_yaml() for the entry point.

apply_env_vars() → None

Copy the env block into os.environ, skipping null values.

Called by from_yaml(); exposed so a config built in code can do the same. Keys with a None value are left alone rather than cleared.

classmethod from_yaml(path: str | pathlib.Path) → CONFIG

Load and validate a YAML file into a CONFIG.

Also records file_path and applies the env block to os.environ – so loading a config has a side effect on the process.

Parameters:

path – path to the YAML file.

Returns:

the validated config.

cuda: CudaConfig | None = None

custom_training_args: CustomTrainingArgsConfig | None = None

data_collator: DataCollatorConfig | None = None

ds: DatasetConfig | None = None

env: dict[str, str | None] | None = None

eval: EvalConfig | None = None

file_path: str | None = None

generation_config: _Flex | None = None

mode: micm_nlp.enums.ModeSE

model: ModelConfig | None = None

peft: PeftConfig | None = None

task: TaskConfig | None = None

test: TestConfig | None = None

tokenizer: TokenizerConfig | None = None

trainer: TrainerConfig | None = None

training_args: TrainingArgsConfig | None = None

class micm_nlp.config.CudaConfig

Bases: _Flex

The cuda block. empty_cache_steps frees the allocator cache every N steps, trading a little speed for headroom.

empty_cache_steps: int | None = None

class micm_nlp.config.CustomTrainingArgsConfig

Bases: _Flex

Settings this package adds beyond HuggingFace’s TrainingArguments.

Three groups. Batching: *_force_sequential and *_max_tokens_per_batch – mutually exclusive, since token-budget batching needs length-sorted order that a sequential sampler would override. Early stopping: patience, threshold, a floor before stopping is allowed, and a metric that is deliberately separable from the one used to pick the best checkpoint. Everything else: which columns to keep, per-parameter-group optimizer settings, a generation whitelist, and what to save at the end.

early_stopping_after: float | None = None

early_stopping_metric: str | None = None

early_stopping_patience: int | None = None

early_stopping_threshold: float | None = None

eval_force_sequential: bool = False

eval_max_tokens_per_batch: int | Literal['auto'] | None = None

generation_whitelist: list[str] | None = None

keep_only_final_model: bool = False

optimizer_grouped_parameters: list[_Flex] | None = None

random_task_exclusion: bool = False

save_final_model: bool = True

test_force_sequential: bool = False

test_max_tokens_per_batch: int | Literal['auto'] | None = None

train_force_sequential: bool = False

usable_columns: list[str] | None = None

class micm_nlp.config.DataCollatorConfig

Bases: _Flex

Thin shell. The actual schema lives in HF (DataCollatorForLanguageModeling, DataCollatorForSeq2Seq, etc.) or in our custom collators. cls selects which collator to instantiate; args is splatted into its constructor.

args: _Flex | None = None

cls: str | None = None

class micm_nlp.config.DatasetConfig

Bases: _Flex

The ds block: which dataset, where it lives, and how to read it.

dirs is a path template under artefacts/datasets/<category>; consumer repos substitute into it (a language segment, a fold) to assemble a run’s data.

Y: _Flex | None = None

category: str | None = None

comes_with_splits: SplitsConfig | None = None

descriptive_name: str | None = None

dirs: str | None = None

input: InputConfig | None = None

label: LabelConfig | None = None

name: str | None = None

preproc_rules: _Flex | None = None

task_id: TaskIdConfig | None = None

type: str | None = None

class micm_nlp.config.EvalConfig

Bases: _Flex

The eval block: when evaluation runs.

Before and after training, on the validation split or the test split, and during training. per_task groups metrics by task id.

after_training: bool = False

after_training_on_test: bool = False

before_training: bool = False

before_training_on_test: bool = False

downstream_tasks: bool | _Flex = False

during_training: bool | _Flex | None = None

per_task: _Flex | None = None

class micm_nlp.config.InitConfig

Bases: _Flex

TRAIN-from-scratch spec: which model class to instantiate and which HF config to pass it. Both cls fields fall back to arch-derived defaults when omitted.

cls: str | None = None

config: InitModelConfigConfig | None = None

class micm_nlp.config.InitModelConfigConfig

Bases: _Flex

HF Config class + its constructor kwargs. Used in TRAIN mode to build the model-config object (e.g. BertConfig) from scratch.

args: _Flex | None = None

cls: str | None = None

class micm_nlp.config.InputConfig

Bases: _Flex

Which dataset column(s) hold the input.

key_2 and key_3 cover pair and triple inputs (premise/hypothesis, context/question/answer). standardize_key renames the column to the canonical name instead of carrying the original through.

key: str

key_2: str | None = None

key_3: str | None = None

standardize_key: bool = False

class micm_nlp.config.LabelConfig

Bases: _Flex

Which column holds the label, and what the label space is.

names and number must agree when the config asks for id-to-name mapping; CONFIG validates that. padded is the id used to pad label sequences, which the loss ignores.

key: str

names: list[str] | None = None

number: int | None = None

padded: int | None = None

standardize_key: bool = False

class micm_nlp.config.ModelConfig

Bases: _Flex

The model block: architecture, and how the model is obtained.

Exactly one of init (build from scratch) or pretrained (load) is used, decided by mode. architecture is a free-form string used for run-directory naming – deliberately not validated against ModelArchSE. The param_size fields are filled in at runtime, not by YAML.

architecture: str

init: InitConfig | None = None

param_size: str | None = None

pretrained: PretrainedConfig | None = None

trainable_param_size: str | None = None

trainable_param_size_ratio: str | None = None

uuid4: str | None = None

class micm_nlp.config.PeftConfig

Bases: _Flex

The peft block: which PEFT method, and how it is parameterized.

A superset of the fields the supported methods need, so peft_type: LORA and peft_type: XPE share one schema; unused fields stay None. The encoder_* fields belong to the Cross-Prompt Encoder – notably encoder_ratio, which is what separates SPT (0), DUAL (between) and XPE (1).

encoder_dropout: float | None = None

encoder_embedding_freeze: bool = False

encoder_embedding_init_type: str = 'hf_default'

encoder_embedding_normalize: str | None = None

encoder_embedding_normalize_max_norm: float | None = None

encoder_freeze: bool = False

encoder_hidden_size: int | None = None

encoder_init_state_dict_path: str | None = None

encoder_input_size: int | None = None

encoder_num_layers: int | None = None

encoder_ratio: float | None = None

encoder_reparameterization_type: str | None = None

num_tasks: int | None = None

num_virtual_tokens: int | None = None

peft_type: str | None = None

task_type: str | None = None

class micm_nlp.config.PostprocConfig

Bases: _Flex

task.preproc_rules: what happens to predictions before metrics see them.

Each flag is a step – flatten, drop padded positions, decode ids to text, map label ids to names, strip and lowercase, coerce to float or back to ids – applied in the order evals.eval runs them. label_restricted_likelihood is the opt-in that scores only the candidate label tokens rather than the whole vocabulary.

calc_confusion_matrix: bool = False

decode: bool = False

filter_by_prefixes: bool | list[str] = False

filter_padded: bool = False

flatten: bool = False

label_id_to_name: bool = False

label_name_strip_lower: bool = False

label_name_to_float: bool = False

label_name_to_id: bool = False

label_restricted_likelihood: bool = False

prediction_axis: int = -1

verify_labels_match: bool = False

class micm_nlp.config.PretrainedConfig

Bases: _Flex

Locates a pretrained model, and names the class to load it with.

cls is resolved by name at runtime, so a new backbone usually needs no code change. An adapter here loads PEFT weights on top of the base model.

adapter: AdapterConfig | None = None

args: _Flex | None = None

checkpoint: str | None = None

cls: str | None = None

name: str | None = None

source: str | None = None

uuid4: str | None = None

class micm_nlp.config.SplitsConfig

Bases: _Flex

Which splits a dataset already ships with.

Each field is False when absent, or the split’s name when present – a string because the on-disk name is not always train/test/validation.

test: bool | str = False

train: bool | str = False

validation: bool | str = False

class micm_nlp.config.TaskConfig

Bases: _Flex

The task block: what is being learned and how it is scored.

metric_groups is a list because one run can score several tasks separately; preproc_rules is the post-processing chain applied before scoring.

category: str | None = None

id: str | None = None

metric_groups: list[_Flex] | None = None

name: str | None = None

preproc_rules: PostprocConfig | None = None

class micm_nlp.config.TaskIdConfig

Bases: _Flex

Which column identifies the task, for multi-task runs that score each task separately.

key: str

standardize_key: bool = False

class micm_nlp.config.TestConfig

Bases: _Flex

The test block: whether and how the held-out test split is scored.

zero_shot adds an untrained baseline pass; zero_shot_only skips training altogether, which is how a zero-shot row is produced.

report_to_wandb: bool = False

run: bool = False

save_predictions: bool = False

zero_shot: bool = False

zero_shot_only: bool = False

class micm_nlp.config.TokenizerConfig

Bases: _Flex

The tokenizer block: which tokenizer to load, or which to train.

adapt_to_lm applies the target architecture’s special tokens and post-processor to a tokenizer borrowed from elsewhere.

adapt_to_lm: bool = False

algorithm: str | None = None

name: str | None = None

source: str | None = None

type: str | None = None

vocab_size: int | None = None

class micm_nlp.config.TrainerConfig

Bases: _Flex

Thin shell. cls selects which HF Trainer subclass to instantiate (Trainer, Seq2SeqTrainer, …). args is reserved for future extra kwargs splatted into the trainer ctor (runtime wiring currently fills the rest).

args: _Flex | None = None

cls: str | None = None

class micm_nlp.config.TrainingArgsConfig

Bases: _Flex

Thin shell. The actual schema lives in HF (TrainingArguments, Seq2SeqTrainingArguments, etc.). cls selects which HF dataclass to instantiate; args is splatted into its constructor at runtime.

For tokenizer-training configs, cls may name a non-HF trainer (e.g. SentencePieceTrainer) and args carries that trainer’s kwargs.

args: _Flex | None = None

cls: str | None = None