Replaces the title-only v3 streams. Same authors, same split assignment (byte-identical to the old split_v2.json), same chunk boundaries -- the packer measures the CONDITION rendering, which does not carry key items. Only the target prompt changes:
Generate an abstract for the title "{title}" using the following items: {clause}
Key items regenerated for all 1,701,966 papers with LongLaMP's verbatim prompt (arXiv 2407.11016), Qwen2.5-7B-Instruct bf16 greedy, vLLM 0.29.0. Validated against the 10,783 papers carrying LongLaMP's real clause: parse rate 0.9985, keyword words 2.35/2/4 vs gold 2.31/2/4.
Removed: authors_v1.parquet (superseded by authors_v2, which is v1 + key_items), split_v2.json (superseded by the identical split_v3.json).
Added: key_items-*-of-006.parquet, key_items_gold_pred, key_items_gold_eval.json, key_items.verify.json.
README corrected: the held-out train-overlap is 57.7%% (valid) / 61.0%% (test), not ~99%% -- the old figure was the corpus-wide seen_in_train, whi…
