localization of LocalLLaMA/typed-decisions, source revision c9e8c4167bf9b47edfcba3b9580e69bddd8d9e5a. Credit: LocalLLaMA and the original dataset contributors. License: Apache-2.0.

Translated and then edited for fluency against the source with RedHatAI/gemma-4-26B-A4B-it-FP8-dynamic at revision ed35d7abe5d940da41b4ff06eb482feb0be8cb44 using vLLM. Shared question and rubric templates also receive an assistant editorial review for terminology and phrasing; this is not human native-speaker review. All five configurations and both train/test splits are retained. The all configuration repeats the same 1,600 cases from the four workflows; it is not additional training data.

Natural-language state text, question instructions, and criterion descriptions are translated. Categorical codes, JSON keys, role labels, IDs, amounts, dates, gold distributions, flat answer columns, factors, and label agreement remain unchanged. state and questions remain JSON-encoded strings with their original nested shapes. Each Parquet file retains its original Arrow schema and row order.

This is a machine translation. Structural validation does not establish semantic equivalence or recalibrate the English teacher probabilities. Original leaderboard scores do not apply to this localization. The test split remains separate from training.

Downloads last month: 81

Total file size:

1.48 MB