HELM Instruct scenario that expands a CFG of Gridfiti ChatGPT prompts and scores free-form replies with a 1-5 Helpfulness critique.
unassessed
| Category | instruction-following |
|---|---|
| Subcategory | HELM Instruct CFG expansion of Gridfiti ChatGPT prompts, scored by Helpfulness |
| Page status | unknown |
| Metric | Helpfulness |
| Direction | higher_is_better |
| Unit | 1-5 |
| Dataset size | 320 |
| Dataset licence | Apache-2.0 |
| Publisher | Stanford CRFM (HELM Instruct); prompt list attributed to Gridfiti Staff |
grammar, as this id, is not linguistic acceptability and not [glue_cola](glue_cola.md) or [blimp](blimp.md). It is Stanford HELM Instruct's GrammarScenario: a context-free grammar expands into English user prompts, the model writes a free-form reply, and a critique metric rates Helpfulness. The checked-in grammar best_chatgpt_prompts.yaml restates GRIDFITI's 2023 "best ChatGPT prompts" list, with slots such as language, city, and event. The HELM schema file says the group should have been named best_chatgpt_prompts, but results use grammar.
Zero-shot generation via get_instruct_adapter_spec (max_tokens 512, temperature 0.7, max_train_instances 0). References are empty. Run spec grammar:path=<yaml>,tags=<csv> keeps only derivations whose collected tags include every requested tag. Scoring is InstructionFollowingCritiqueMetric.
No model card in ModelSpec reports this benchmark yet.