---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: question
    dtype: string
  - name: choices_text
    sequence: string
  - name: choices_label
    sequence: string
  - name: instruction
    dtype: string
  - name: ground_truth
    dtype: string
  - name: has_label
    dtype: bool
  - name: fact1
    dtype: string
  - name: fact2
    dtype: string
  - name: combinedfact
    dtype: string
  - name: domain
    dtype: string
  - name: dataset
    dtype: string
  - name: scoring_kind
    dtype: string
  - name: num_choices
    dtype: int32
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

# QASC formatted for MAE / SFT evaluation

This dataset mirrors `allenai/qasc` with columns kept convenient for local
inference and scoring.

- `instruction` is the formatted multiple-choice prompt used by this repo.
- `ground_truth` is the answer letter when available.
- `has_label=false` on the official blind test split, whose `answerKey` is not public.
- `id` is preserved for joining with inference outputs.