Answer Table Question (HuggingFace)

How it fits

Typical backends

Caveats

• I/O contractThe first input is the table as a JSON String and the second is the question String. The table JSON may be a column-oriented dict (`{"Actors": ["Brad Pitt", ...]}`), a `{"columns": [...], "rows": [[...]]}` header-plus-matrix, or a list of row objects; every shape is read into a pandas DataFrame and stringified for the tokenizer.
• Parameter interactionTAPAS predicts both the answer cells and an aggregation operator. NONE joins the selected cell text, COUNT returns the number of selected cells, and SUM / AVERAGE parse the selected cells as numbers and reduce them, so one TAPAS pass answers both lookup and numeric questions.
• FallbackAn empty question, an empty or column-less table, or an unparseable table JSON short-circuits to an empty String answer, and an inference failure on a single message is logged and also returns an empty answer rather than crashing the stream; every input emits exactly one answer.
• Hard constraintTAPAS caps the flattened question-plus-table at 512 tokens; larger tables are truncated row by row to fit, so answers depend only on the rows that survive truncation.
• Latencymodel is hot-swappable at runtime and triggers a reload (and a GPU memory reclaim on CUDA) on the next call, producing a one-call latency spike. device is captured once at startup; TAPAS-base is small enough to run on CPU.