pretty_name: BIG-Bench Hard Freeform Rewrites
language:
- en
configs:
- config_name: boolean_expressions
data_files:
- split: train
path: boolean_expressions/train-*
- config_name: causal_judgement
data_files:
- split: train
path: causal_judgement/train-*
- config_name: date_understanding
data_files:
- split: train
path: date_understanding/train-*
- config_name: disambiguation_qa
data_files:
- split: train
path: disambiguation_qa/train-*
- config_name: dyck_languages
data_files:
- split: train
path: dyck_languages/train-*
- config_name: formal_fallacies
data_files:
- split: train
path: formal_fallacies/train-*
- config_name: geometric_shapes
data_files:
- split: train
path: geometric_shapes/train-*
- config_name: hyperbaton
data_files:
- split: train
path: hyperbaton/train-*
- config_name: logical_deduction_five_objects
data_files:
- split: train
path: logical_deduction_five_objects/train-*
- config_name: logical_deduction_seven_objects
data_files:
- split: train
path: logical_deduction_seven_objects/train-*
- config_name: logical_deduction_three_objects
data_files:
- split: train
path: logical_deduction_three_objects/train-*
- config_name: movie_recommendation
data_files:
- split: train
path: movie_recommendation/train-*
- config_name: multistep_arithmetic_two
data_files:
- split: train
path: multistep_arithmetic_two/train-*
- config_name: navigate
data_files:
- split: train
path: navigate/train-*
- config_name: object_counting
data_files:
- split: train
path: object_counting/train-*
- config_name: penguins_in_a_table
data_files:
- split: train
path: penguins_in_a_table/train-*
- config_name: reasoning_about_colored_objects
data_files:
- split: train
path: reasoning_about_colored_objects/train-*
- config_name: ruin_names
data_files:
- split: train
path: ruin_names/train-*
- config_name: salient_translation_error_detection
data_files:
- split: train
path: salient_translation_error_detection/train-*
- config_name: snarks
data_files:
- split: train
path: snarks/train-*
- config_name: sports_understanding
data_files:
- split: train
path: sports_understanding/train-*
- config_name: temporal_sequences
data_files:
- split: train
path: temporal_sequences/train-*
- config_name: tracking_shuffled_objects_five_objects
data_files:
- split: train
path: tracking_shuffled_objects_five_objects/train-*
- config_name: tracking_shuffled_objects_seven_objects
data_files:
- split: train
path: tracking_shuffled_objects_seven_objects/train-*
- config_name: tracking_shuffled_objects_three_objects
data_files:
- split: train
path: tracking_shuffled_objects_three_objects/train-*
- config_name: web_of_lies
data_files:
- split: train
path: web_of_lies/train-*
- config_name: word_sorting
data_files:
- split: train
path: word_sorting/train-*
BIG-Bench Hard Freeform Rewrites
DatologyAI is pleased to release a rewrite of BIG-Bench Hard that enables language-model evaluation through language-modeling loss on verified-correct free-form answers. Part of the DatologyAI Benchmark Freeform Rewrites collection.
Motivation
Practitioners developing language models need evaluations capable of quantifying the effects of small changes, such as adjustments to the training data mix, even when evaluating the smaller models trained during research and development.
Generative evaluation with pass/fail checks is susceptible to emergence effects when applied to development-scale models. Unconditioned language-modeling loss over held-out documents provides a denser signal, but held-out documents may not map clearly onto the capabilities of interest.
Few-shot-conditioned free-form question answering, the setup this dataset enables, sits between the two. It targets specific capabilities and measures in-context learning while providing a dense, emergence-free signal.
This data
The DatologyAI Benchmark Freeform Rewrites datasets transform existing multiple-choice and short-answer benchmarks into free-form long-answer tasks. Each row pairs a self-contained question, the original short answer, and a worked continuation verified to answer the question correctly.
A technical write-up describing the rewrite and verification process and the validation results is coming soon.
Fields
| Field | Description |
|---|---|
id |
<group>#<source id>; unique within the dataset. |
group |
Subject or subtask; equals the HF config name. |
question |
Self-contained rewritten question or problem statement. |
answer |
Short original answer from the source benchmark. |
freeform_answer |
Free-form worked answer that arrives at answer. |
structural_type |
How the source multiple-choice item was converted: self_contained_factoid, fill_in_blank, multi_statement, negative_odd_one_out, logic_symbolization or competing_proposition. |
Loading
from datasets import load_dataset
ds = load_dataset("DatologyAI/bbh-freeform-rewrites", "boolean_expressions", split="train")