A parallel multilingual reasoning dataset in which every problem is solved through a fixed five-stage reasoning scaffold rather than a single free-form explanation. Each record restates the goal, lays out a solution plan, establishes the premises and variables it will rely on, works through exactly five numbered reasoning steps, and closes with a consolidated answer. The same problem is carried across eight languages with all five stages aligned, so a model's reasoning can be compared step for step across language boundaries instead of only at the final answer.
Questions are written in the register of a real user prompt — conversational and open-ended in the social and economic domains, precisely specified in the mathematical and engineering ones — so the dataset trains reasoning on the kind of input a deployed model actually receives.
| Stage | Content |
|---|---|
question | Given yz − ln(z) = x + y, where z is a function of x and y, find ∂z/∂x. |
task | Express ∂z/∂x explicitly under the differentiability condition. |
approach | Differentiate both sides with respect to x, applying the chain rule to z. |
initial_analysis | y is independent, so it is constant under ∂/∂x; the chain rule introduces ∂z/∂x. |
step_by_step | Five numbered steps: differentiate each side → factor (y − 1/z)·∂z/∂x = 1 → solve → verify. |
output | ∂z/∂x = z / (yz − 1) |
step_by_step segments per locale; LaTeX in $$...$$ preserved verbatim across languagesProblems are authored per domain across Social & Culture, Economy, Math, Science, and Engineering, and assigned a domain-namespaced sid so that a problem stays traceable across every language version. Questions are written in the register in which a user would actually ask them, which keeps conversational and underspecified prompts in the set rather than restricting it to exam-style phrasing.
Each problem is then localized into the eight target languages. Localization covers all six fields, not the question alone, so that the plan, premises, and every reasoning step exist in parallel. Notation, variable names, and LaTeX expressions are carried over unchanged so that mathematical content is identical across languages while the surrounding prose reads naturally to a native reader.
Reasoning is written into the fixed five-stage scaffold — restate the goal, state the plan, fix the premises, work through five numbered steps, consolidate the answer. Each stage has a defined role: the plan may not contain results, the premises may not introduce conclusions, and the closing answer may not add facts absent from the steps.
Reviewers check each record on four axes. First, stage-role integrity: whether each field does its own job rather than restating an adjacent one. Second, the five-step invariant: whether the reasoning genuinely divides into five substantive steps rather than being padded or compressed to reach the count. Third, cross-locale parity: whether all eight language versions carry the same step count, the same reasoning path, and the same final answer. Fourth, entailment: whether the closing answer follows from the steps as written, with no unsupported addition.
For mathematical and engineering records, reviewers additionally re-derive the result independently and verify that LaTeX expressions and technical terms survived localization intact. Records failing any axis are rewritten rather than patched, since a broken stage boundary cannot be repaired by editing a single sentence.