multi-turn123

This dataset repo contains four transformed long-context / multi-turn evaluation configs:

• niah: grouped sequence format, 3 long contexts per sample.
• qasper: grouped sequence format, 3 papers per sample.
• qmsum: grouped sequence format, 3 transcripts per sample.
• prefeval: explicit PrefEval static-history samples with 300 history turns, the preference disclosure inserted at the 30th turn from the end (turn 271), and the final question kept only in queries[0].

All configs use a single test split.

Load

from datasets import load_dataset

niah = load_dataset("ezh18/multi-turn123", "niah", split="test")
qasper = load_dataset("ezh18/multi-turn123", "qasper", split="test")
qmsum = load_dataset("ezh18/multi-turn123", "qmsum", split="test")
prefeval = load_dataset("ezh18/multi-turn123", "prefeval", split="test")

Sequence Datasets

niah, qasper, and qmsum are event-stream style records. Each sample has a sequence list. When item_type == "transcript", insert the long text into the running context. When item_type == "question", ask that question against the context accumulated so far. For each non-final long text in a 3-item group, the last two questions are randomly shifted to a position after the next long-text item.

PrefEval Static Variant

prefeval is the PrefEval-static-300-history-last30-IVF data variant. Each row contains:

• context: serialized history without the final question.
• queries: a one-item list containing the final PrefEval question.
• answers: a one-item list containing the target explicit preference.
• labels: judge/debug metadata, including the target preference and final question.
• segment_meta: role-prefixed history segment spans; the explicit preference segment has is_preference_segment=True and turn_index=271.
• num_history_turns: 300 for every sample.
• pref_insert_turn: 271 for every sample, leaving 29 subsequent history turns before the final question.

This dataset is intended for one-shot static prefill + one-shot IVF index construction + final-query evaluation. It is not meant to test online per-turn IVF cache rebuilding.