TAT-QA

What it measures

Financial questions combine evidence from tables and accompanying prose, often requiring arithmetic or multiple answer spans.

How to read the result

Preserve units and scale, such as thousands, millions and percentages. Span extraction and numerical reasoning expose different errors.

Limits and comparison traps

Use the official answer normalization and task split. Full-context question answering and retrieval-augmented answering are different evaluation setups.

Published evaluation

Published test-set result over hybrid financial tables and text. Exact match and token F1 remain distinct.

Where this fits in an evaluation suite

Synthetic research rewards do not establish live investment performance. Preserve transaction costs, held-out regimes, constraints, and the exact policy behind a reported result.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

Published test-set result over hybrid financial tables and text. Exact match and token F1 remain distinct.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.