HELM Enterprise wrap of ConvFinQA: given a table, gold text facts, and prior turns, emit the last-turn number and score float equality.
unassessed
| Category | domain |
|---|---|
| Subcategory | HELM Enterprise numeric last-turn answers on ConvFinQA tables |
| Page status | unknown |
| Metric | float_equiv |
| Direction | higher_is_better |
| Dataset size | 12594 |
| Dataset licence | MIT |
| Publisher | Stanford CRFM (HELM scenario); ConvFinQA data from UC Santa Barbara and J.P. Morgan |
conv_fin_qa_calc is Stanford HELM's calculation scenario built on ConvFinQA (Chen et al., EMNLP 2022). Each instance is one conversation turn: a markdown table from an earnings report, optional gold supporting-fact sentences, earlier questions with their gold numeric answers, and the last unanswered question. The model must output that last number. It does not ask for ConvFinQA's reasoning program. English text plus a table. HELM Enterprise run spec, not the HELM Finance fin_qa program-generation task.
Generation. Instructions: "Based on the table, answer the final question. Respond with the answer only, with no additional explanation." Output noun Answer. Default adapter max_tokens 5 and max_train_instances 5 (five-shot from TRAIN_SPLIT). Main split valid. Main metric float_equiv.
No model card in ModelSpec reports this benchmark yet.