Safety evidence map of AI for mental health: coded dataset, codebook and scripts

Coded dataset, codebook and analysis scripts for the paper 'Coverage and Predictors of Safety Testing in Studies of Artificial Intelligence for Mental Health: Scoping Review'. 1,986 unique records from 2,042 open-access PDFs on LLMs, chatbots, AI therapy and AI safety, coded by large language models for nine safety dimensions; includes the full-text check, retraction checks (with and without DOI), two recall checks, a cross-family coder check (all 997 coded records recoded twice with Google Gemini 4 Argon, OpenAI GPT-6.1 Sol and DeepSeek V4 Pro) and a PubMed database search (4 November 2026; 2,963 hits listed by PMID, matched to the corpus with a manual review of title matches, and coded from title and abstract). Full texts and PubMed abstracts are not redistributed. See README.md for limits. Version 1.4 corrects the matching of PubMed hits to the corpus (hits matched to removed duplicate copies; 15 false title matches and one false DOI match, reviewed by hand in pubmed/match_review.csv), so the PubMed coverage is 28% to 31% instead of 26% to 30%, and adds the second coder's codes, the repeat-record codes and a script that recomputes the reliability statistics from the deposit. Version 1.3 added the PubMed search (folder pubmed/), a clean second cross-family run with per-record settings (crossfamily/run2/), the retraction check for the 79 included studies without a DOI, and scorers for a human check. See What changed in version 1.4 in README.md.