Note Recommended benchmark entry. Use `xhotpotqa_v1_1_audited`: 15,661 train + 7,405 validation rows, per-row audit status, and recovered original English `source_sentences`. The earlier V1 config remains available for reproducibility.

Note Audited V2 RC1 generated with Gemma 4 31B through a vLLM-served OpenAI-compatible endpoint. Covers 22,836/23,066 intended sources; 230 missing IDs and incomplete assignment provenance prevent canonical status.

Note Sanitized, balanced source-referenced GLM-5.2 translation-quality audit for V1: 2,760 scored units across 23 target languages and paragraph/question/answer strata.

Note Sanitized, balanced source-referenced GLM-5.2 translation-quality audit for V2 RC1: 2,760 valid units. The 240 superseded mixed-run answer rows are excluded by construction.