crisis-sitrep-eval

Evaluation resources for comparing LLM-generated and professional humanitarian situation reports (sitreps), from "Looking Complete Is Not Being Grounded: Evaluating LLM Sitreps Against Human-Authored Situation Reports" (LT4CPR Workshop @ AACL-IJCNLP 2026).

• Paper: https://arxiv.org/abs/2610.18762
• Code: https://github.com/techsachinkr/llm-sitrep-evaluation

Contents

File	What it holds
frozen_schema.yaml	The 15-slot information schema, with definitions and two synthetic examples per slot
per_judgement.csv (judgements)	Slot-level judge verdicts (ABSENT / PARTIAL / PRESENT) for 121 professional and 261 LLM-generated reports across five disasters, without text
tables/human_validation_pairs.csv (human_validation)	100 blind human labels with the four judges' verdicts (Cyclone Idai)
tables/	Every table behind the paper's numbers: completeness, slot coverage, figure grounding, source availability, metric correlations, sensitivity analyses
reliefweb_report_manifest.csv	Titles, organisations, dates, URLs and hashes of the 125 ReliefWeb reports
humaid_source_manifest.csv	HumAID tweet IDs and text hashes (76,484 rows)
machine_report_manifest.csv	Request hash, output hash, call time and model ID for each of the 264 generations
prompts.json, measurement_spec.json, model_roles.csv, decoding_parameters.csv	Prompts, measurement rules, models and decoding settings
schema_audit/open_codes.csv (open_codes)	All 1,360 open codes by report (name, definition, confidence, cluster), without evidence spans
schema_audit/clusters_summary.csv (clusters), schema_audit/clusters.csv	All 382 clusters (threshold .35), ranked, with the 25 candidates flagged and every member code
schema_audit/ (other files)	Induction sample, the 25 candidates, merge decisions and rationales, review record, frozen schema, and the embedding cache used for clustering
REPRODUCIBILITY.md	How to re-run the analysis from these files

What is not included

No evidence spans from the open codes, no tweet text, no generated report text (generations can repeat personal data from crisis posts), and no ReliefWeb documents (copyright of the publishing organisations). Use the manifests to retrieve the sources and verify them against the hashes.

Limitations

The schema was induced from 56 Cyclone Idai reports, 44 of them OCHA's, so it chiefly reflects OCHA reporting practice. The human validation used one annotator on Idai only. Judge labels are LLM outputs. HumAID is a labelled sample of English-language tweets, not social media at large.

Licence

CC BY-NC-SA 4.0, inherited from HumAID (Alam et al., 2021) for the tweet-ID manifest.

Citation

@misc{kumar2026sitrepgap,
  title         = {Looking Complete Is Not Being Grounded: Evaluating {LLM} Sitreps Against Human-Authored Situation Reports},
  author        = {Kumar, Sachin},
  year          = {2026},
  eprint        = {2610.18762},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2610.18762}
}