Skip to content

Question Answering

Bases: TextFineTuner

A bolt for fine-tuning Hugging Face models on question answering tasks.

Parameters:

Name Type Description Default
input BatchInput

The batch input data.

 required
output OutputConfig

The output data.

 required
state State

The state manager.

 required

CLI Usage:

    genius QAFineTuner rise \
        batch \
            --input_folder ./input \
        batch \
            --output_folder ./output \
        none \
        --id microsoft/tapex-large-finetuned-wtq-lol \
        fine_tune \
            --args \
                model_name=my_model \
                tokenizer_name=my_tokenizer \
                num_train_epochs=3 \
                per_device_train_batch_size=8

__init__(input, output, state, **kwargs)

Initialize the bolt.

Args:
    input (BatchInput): The batch input data.
    output (OutputConfig): The output data.
    state (State): The state manager.
    **kwargs: Additional keyword arguments.

compute_metrics(eval_pred)

Compute the accuracy of the model's predictions.

Parameters:

Name Type Description Default
eval_pred tuple

A tuple containing two elements: - predictions (np.ndarray): The model's predictions. - label_ids (np.ndarray): The true labels.

 required

Returns:

Name Type Description
dict Optional[Dict[str, float]]

A dictionary mapping metric names to computed values.

load_dataset(dataset_path, pad_on_right=True, max_length=None, doc_stride=None, evaluate_squadv2=False, **kwargs)

Load a dataset from a directory.

Supported Data Formats and Structures:

JSONL

Each line is a JSON object representing an example. 

{"context": "The context content", "question": "The question", "answers": {"answer_start": [int], "text": [str]}}

CSV

Should contain 'context', 'question', and 'answers' columns. 

context,question,answers
"The context content","The question","{'answer_start': [int], 'text': [str]}"

Parquet

Should contain 'context', 'question', and 'answers' columns.

JSON

An array of dictionaries with 'context', 'question', and 'answers' keys. 

[{"context": "The context content", "question": "The question", "answers": {"answer_start": [int], "text": [str]}}]

XML

Each 'record' element should contain 'context', 'question', and 'answers' child elements. 

<record>
    <context>The context content</context>
    <question>The question</question>
    <answers answer_start="int" text="str"></answers>
</record>

YAML

Each document should be a dictionary with 'context', 'question', and 'answers' keys. 

- context: "The context content"
  question: "The question"
  answers:
    answer_start: [int]
    text: [str]

TSV

Should contain 'context', 'question', and 'answers' columns separated by tabs.

Excel (.xls, .xlsx)

Should contain 'context', 'question', and 'answers' columns.

SQLite (.db)

Should contain a table with 'context', 'question', and 'answers' columns.

Feather

Should contain 'context', 'question', and 'answers' columns.

Parameters:

Name Type Description Default
dataset_path str

The path to the dataset directory.

 required
pad_on_right bool

Whether to pad on the right.

 True
max_length int

The maximum length of the sequences.

 None
doc_stride int

The document stride.

 None
evaluate_squadv2 bool

Whether to evaluate using SQuAD v2 metrics.

 False

Returns:

Name Type Description
Dataset Optional[Dataset]

The loaded dataset.

prepare_train_features(examples, cls_token_id=None)

Tokenize our examples with truncation and padding, but keep the overflows using a stride.

Parameters:

Name Type Description Default
examples Dict[str, Union[str, List[str]]]

The examples to be tokenized.

 required

Returns:

Type Description
Optional[Dict[str, Union[List[int], List[List[int]]]]]

The tokenized examples.