pretty_name: DROP Freeform Rewrites
language:
- en
configs:
- config_name: drop
data_files:
- split: train
path: drop/train-*
DROP Freeform Rewrites
DatologyAI is pleased to release a rewrite of DROP 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 |
Item id from the source benchmark; unique within the dataset. |
question |
Self-contained rewritten question or problem statement. |
passage |
Grounding passage the question is asked about. |
answer |
Short original answer from the source benchmark. |
freeform_answer |
Free-form worked answer that arrives at answer. |
Loading
from datasets import load_dataset
ds = load_dataset("DatologyAI/drop-freeform-rewrites", split="train")