An OpenCompass generation config over a local qabench-test.qa.csv of prompt/reference pairs; the public tree does not identify a paper, licence, or item count.
unassessed
| Category | knowledge |
|---|---|
| Subcategory | OpenCompass local-CSV question answering (prompt/reference pairs; origin not established) |
| Page status | unknown |
| Direction | higher_is_better |
| Publisher | OpenCompass (config host); original dataset publisher not established |
OpenCompass qabench is a generation task that feeds each row's `prompt` field to the model and compares the reply with a `reference` string. The shipped config uses HuggingFace `load_dataset` on a local CSV at `./data/qabench/qabench-test.qa.csv`, loaded as split `train` even though the filename says test. No README in the config directory describes the questions, the language, or how the references were written. Hugging Face dataset search for "qabench" returned an empty list on 2026-09-08. This page documents that OpenCompass task, not Q-Bench (low-level vision) or other QA-Bench names.
Zero-shot free-text generation. The human turn is the raw `{prompt}` string. The config names no shot count and no option letters.
No model card in ModelSpec reports this benchmark yet.