OpenCompass dataset id dingo: generate from local English/Chinese CSVs, then score completions with dingo-python llm_base rules.
unassessed
| Category | generation |
|---|---|
| Subcategory | OpenCompass generation + dingo-python llm_base rules on English and Chinese prompts |
| Page status | unknown |
| Metric | dingo-python Executor summary (eval_group=llm_base) |
| Direction | higher_is_better |
| Dataset licence | Apache-2.0 |
| Publisher | OpenCompass (harness wrap); MigoXLab / DataEval (dingo-python) |
OpenCompass configs/datasets/dingo/dingo_gen.py defines two zero-shot generation runs whose predictions are scored by DingoEvaluator. The prompts come from local files ./data/dingo/en_192.csv and ./data/dingo/zh_170.csv (semicolon-delimited first column). DingoEvaluator writes prompt/prediction jsonl and runs dingo-python's local Executor with eval_group llm_base, then returns that summary dict as the OpenCompass score. This is a wrap of the MigoXLab/DataEval dingo-python data-quality toolkit applied to model outputs, not a held-out QA accuracy set.
OpenCompass GenInferencer, ZeroRetriever, PromptTemplate with a single HUMAN round equal to {input}. No gold output_column. Evaluator pred_role is BOT. A second loader, DingoLongDataset, reads jsonl {"input": ...} but is not referenced by dingo_gen.py.
No model card in ModelSpec reports this benchmark yet.