q-topic-finetuning_skill

Overview

This skill fine-tunes topic-model outputs into a reproducible, theory-driven classification framework tailored for academic manuscripts. It consolidates model topics (BERTopic, LDA, NMF) into final codes and labels while preserving domain-sensitive distinctions and multi-category assignments. The output includes an implementation plan and updated Excel/CSV files with final labels and verification counts.

How this skill works

The skill ingests topic model exports (Excel/CSV) and analyzes topic overlaps, unassigned topics, and counts. You define a FINAL_TOPICS mapping that groups source topic IDs into theory-aligned codes and themes. It generates assignment mappings, calculates overlap reconciliations, marks multi-category topics, and writes updated classification columns back to the document. It also produces an implementation plan (Markdown) and verification checklist for reproducibility.

When to use it

• Converting raw topic model output into manuscript-ready thematic categories
• Applying a theoretical framework (e.g., legitimacy, stakeholder theory) to topic clusters
• Consolidating many model topics into a smaller, interpretable set
• Preserving entity-, event-, geography-, or stakeholder-specific distinctions
• Updating Excel/CSV source files with final labels and reconciliation counts

Best practices

• Explicitly declare preservation rules for items that must never be merged (entities, events, geographies, stakeholders).
• Build FINAL_TOPICS as a dictionary with label, theme, and source IDs to keep mapping transparent and reproducible.
• Track multi-category assignments in an assignments dictionary and represent themes as semicolon-separated values for clarity.
• Compute overlap totals and verify: non-outlier_docs + total_overlap = original total to reconcile counts.
• Keep a small implementation plan and verification checklist to ensure reviewers can reproduce merges and counts.

Example use cases

• Merge 50+ BERTopic clusters into 25 theory-driven topics for a literature review manuscript.
• Apply a legitimacy framework to classify topics into cognitive, pragmatic, and moral categories.
• Preserve company-specific topics while consolidating industry-wide themes for comparative studies.
• Mark topics that span multiple categories and export semicolon-separated themes for downstream analysis.
• Update a document-level Excel with Final_Topic_Code, Final_Topic_Label, and Category_Theme columns for reporting.