KLEVER model – Keywords & Labels, LLM Evaluation, Verified by Expert Review

KLEVER model – Keywords & Labels, LLM Evaluation, Verified by Expert Review. KLEVER is a three-stage LLM-guided workflow in which expert-refined keywords isolate relevant records, dual LLMs validate category context, and binary/exemplar-based prompts classify the corpus. Final trigger keywords are clustered and labeled to form child categories, resulting in a structured, curated dataset. Version 2: Datasets Version 3: codes Description: The HumanCrisisData.xlsx dataset is based on the MAPE corpus of 182,491 historical letters from the Arquivo Histórico Ultramarino (Lisbon). The objective is to identify and structurally classify evidence of humanitarian crises across five predefined domains: Natural Disasters & Weather Events, Health Crisis: Epidemics & Diseases, Environmental Degradation, Human Displacement, and Resource Scarcity. The curation process followed a three-stage KLEVER workflow: Keyword-Based Retrieval – LLM-proposed Portuguese crisis keywords were refined by experts and used to extract candidate records from the entire corpus. LLM-as-a-Judge Validation & Context Screening Dual-Gemini/GPT validation checked keyword–category coherence. Binary Classification (Yes/No) was applied to the full corpus. All records labelled 'Yes' were then classified using multi-label context-based assignment (Parent category selection using dynamic few-shot prompt examples). Trigger Word Extraction & Subcategory Structuring – The exact word that caused the classification was identified, vector-clustered, and LLM-labelled into interpretable sub-domains (child categories). The final dataset consists of 8,108 curated records, each containing: archival metadata, original Portuguese text and English translation, multi-label parent category assignment, trigger keyword, and LLM-labelled child category. File Sheet description: The Excel file contains multiple sheets documenting each stage of the KLEVER workflow, from category definition and keyword generation to validation, classification, and final structured outputs. Each sheet represents a progressive step in transforming raw historical correspondence into a curated dataset with assigned crisis categories, child subcategories, and trigger keywords. Category_Description – Defines the scope and meaning of the five parent humanitarian-crisis categories. LLM_EXPERT_Keyword – Initial keyword proposals by LLM refined through expert validation. keyword_found – Records retrieved via keyword-based filtering from the full corpus. keyword_found_summary – Statistical count of keyword occurrences per parent category. Data_Version_1 – Keyword-matched records assigned to parent categories (pre-validation). Data_Version_2 – Dual LLM validation results with keep/discard decisions. SME vs LLM as a Judge – Subject matter expert validation of records Data_Version_3 – Binary Yes/No classification and multi-label parent category assignment. Data_Version_4 – Final structured dataset including parent category, child category, and trigger keyword. trigger_words – Evolution of crisis-related keywords from seed list to final trigger words.