ThakiCloud/korean-contract-derivation-bench

Korean Contract Derivation Benchmark

A model that derives a discount and a unit-price total without being told to will still return the VAT-inclusive figure verbatim when asked for the contract amount — and a generic instruction to "compute the relation" does not fix it. This benchmark isolates that.

The finding

240 synthetic Korean procurement summaries. Half state the contract amount directly among distractor figures (disambiguate); half state it only as a relation the model must compute (derive), in three flavours: VAT-inclusive ÷1.1, list price × 0.8 after a stated 20% discount, and unit price × quantity. Three system prompts, two model arms, one serving run.

Accuracy by derivation type (n=40 per cell):

prompt	disambiguate	discount	unit × qty	VAT
none	1.000	0.83–0.90	0.93–0.98	0.000
generic ("if stated as a relation, compute it")	1.000	0.98–1.00	1.000	0.000
VAT-specific ("divide by 1.1")	1.000	0.95–1.00	0.95–0.98	0.85–0.90

Ranges span the two arms (bf16 and a 4-bit build); both behave the same way.

The model is not bad at deriving. With no instruction at all it handles discount and unit-price arithmetic 83–98% of the time. Only VAT collapses, and it collapses completely. ÷1.1 is not harder than ×0.8 or ×2, so arithmetic is not the explanation.

The generic instruction is worth nothing here. It lifts the cases that already worked and leaves VAT at zero. What we think is happening: 부가세 포함 997,700,000원 ("997,700,000 won including VAT") is such a standard phrase in Korean contracts that the model reads the number as the answer, not as an input to a calculation — so a prompt conditioned on "the value is not stated directly" never fires.

Why this shape of failure is dangerous

It is narrow, total, and invisible next to its neighbours. If your eval set happens to contain only discount-style derivations you will conclude derivation works. We drew exactly that wrong conclusion twice before widening the set from one derivation type to three.

The practical implication for Korean-document RAG: a single generic instruction is not a product asset. What is needed is a lexicon of answer-masquerading phrases per domain, each with its own rule. Which phrases those are differs by document type, so it is a spec item, not prompt tuning.

Contents

• data/cases.jsonl — 240 cases: document (contract summary, padding stripped), question, answer_krw, mode (disambiguate | derive:vat | derive:discount | derive:unit), distractors.
• data/results.jsonl — measured accuracy per (arm, prompt variant, mode).
• data/prompts.json — the three system prompts verbatim.

Every derived figure is exactly invertible from the document (997,700,000 ÷ 1.1 = 907,000,000), and the answer never appears verbatim in a derive document — verified programmatically before the run.

Method

vLLM 0.27.1, greedy decoding, one serving session, identical config across arms, model path the only difference. Scored by string/numeric match against answer_krw. Documents are padded to the target context length by the harness at eval time; the padding is filler and is stripped here.

Limits

• Synthetic. These are generated documents, not real contracts. The distribution of answer-masquerading phrases in actual Korean procurement paperwork is unmeasured.
• The three derivation types are not equally hard — unit × quantity uses small integers and is the easiest. Compare prompt variants within a type, not accuracies across types.
• n=40 per cell. One item is 0.025. Do not read small differences as trends.
• One run, one model family.
• The VAT-specific prompt hands the model the formula. Its 0.85–0.90 shows that instructing works, not that the model knows the rule unprompted.