muc4-chat

MUC-4 document-level event extraction as chat-formatted examples, one per document, in the corpus's train/dev/test splits.

Source. The GTT-preprocessed MUC-4 corpus from xinyadu/gtt (data/muc/processed/; MIT licence for the preprocessing, the MUC-4 data itself is in the public domain).

Example. messages holds a system prompt (identical for every example), a user message with the document text, and an assistant message with the target. n_input_tokens counts the rendered system + user turn with the Qwen/Qwen3-4B-Instruct-2507 tokenizer; truncated marks dev/test documents whose text was cut to fit 1800 input tokens (train documents over budget are dropped instead).

Target. A compact JSON array of the document's events, [] for an irrelevant document. Each event has incident_type (one of attack, bombing, kidnapping, arson, robbery, forced work stoppage) followed by its non-empty roles among PerpInd, PerpOrg, Target, Victim, Weapon, in that order. A role is a list of entities; an entity is the list of its coreferent mentions, verbatim from the document, without character offsets.

Canonical form. Duplicate events (identical once offsets are removed) are collapsed to one. Events are ordered by the offset of their earliest mention (events without mentions last), then by event type, then by their JSON text. Mentions within an entity are ordered by offset, then by string. Empty entities are dropped.

Prepared by src/data/prepare.py in the fine-tuning-decoder project.