README.md

---
dataset_info:
  features:
  - name: source_id
    dtype: string
  - name: source_revision
    dtype: string
  - name: task
    dtype: string
  - name: validator_type
    dtype: string
  - name: validator_params_json
    dtype: string
  - name: ideal
    dtype: string
  - name: answer
    dtype: string
  - name: expected
    dtype: bool
  splits:
  - name: train
    num_bytes: 42198
    num_examples: 39
  download_size: 28278
  dataset_size: 42198
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# SEQQA reward reference

Regression fixture for `trl.internal.seqqa.reward`, used by `tests/internal/test_seqqa_rewards.py`.

39 hand-checked answers across 20 SEQQA validator types: for each the
`answer` string, the `ideal` reference answer, and whether `seqqa_accuracy_reward` is expected to score it
`1.0` (`expected = true`) or `0.0`. `validator_params_json` is the serialised validator payload, and
`source_id` / `source_revision` point at the generated question this row was pinned from.

Values are pinned to real generated questions, so regenerating them is not equivalent to editing this set;
change a row only alongside the reward change that motivates it.