Understanding Large Language Models for Accounting and Finance: Data and Methodology Dr Yuqian Zhang, Auckland University of Technology Date: 29 July 2026 ================================================================================ OVERVIEW ================================================================================ This directory contains the datasets used in the research brief "Understanding Large Language Models and Their Implications for Accounting and Finance Research." Each dataset is described below with its source, compilation method, and known limitations. ================================================================================ DATASET DESCRIPTIONS ================================================================================ 1. llm_parameter_timeline.csv Description: Parameter counts, training compute (FLOP), release dates, and developer information for major large language models from 2017 to 2026. Sources: Confirmed values from original published papers (see CSV header for full citations). Estimated values from industry analysis (SemiAnalysis, March 2023) and scaling-trend projections. Training compute values from Epoch AI database (epoch.ai/trends) for models through 2024; post-2024 values projected at 4-5x per year. Key columns: model, release_year, release_month, parameters_billions, training_compute_flops, developer, notes Limitations: Parameter counts for proprietary models (GPT-4, Claude 3.5, Gemini Ultra, GPT-5, Claude 4.5, Llama 5) have not been confirmed by their developers. Training compute for recent models is estimated from reported infrastructure and training duration. FLOP values are approximate. 2. llm_arena_scores.csv Description: Quarterly LMArena (formerly LMSYS Chatbot Arena) Elo scores for selected frontier models from Q2 2023 to Q2 2026. Sources: LMArena public leaderboard (lmarena.ai). Scores are approximate quarterly means extracted from publicly available leaderboard data. Key columns: model, year_quarter, elo_score Limitations: Elo scores fluctuate with new votes and represent approximate quarterly snapshots. The January 2026 LMArena rebrand shifted some Elo distributions by 20-40 points due to methodology changes (Style Control filtering). Pre- and post-rebrand scores may not be directly comparable. 3. llm_hallucination_rates.csv Description: Hallucination rates on document summarisation benchmarks for major LLMs, 2024 to 2026. Sources: Vectara Hughes Hallucination Evaluation Model (HHEM) Leaderboard (github.com/vectara/hallucination-leaderboard). The HHEM benchmark measures factual inconsistency with a provided source document (intrinsic hallucination). Post-Q3 2025 values are estimated from reported trends and marked accordingly in the data_status column. Key columns: model, date_label, hallucination_rate_pct, data_status Limitations: Rates measure a narrow form of hallucination (consistency with source document). Open-ended generation shows substantially higher rates. These numbers are not transferable across tasks. Estimated values are clearly marked as such. 4. accounting_ai_adoption.csv Description: AI and GenAI adoption rates among accounting and tax firms, from industry surveys conducted in 2024 and 2025. Sources: Wolters Kluwer, "Future Ready Accountant Report" (2024 and 2025 editions, wolterskluwer.com/en/know/future-ready-accountant). Thomson Reuters Institute, "2025 Generative AI in Professional Services Report" (2025). Key columns: source, year, category, percentage Limitations: Survey samples may not represent the entire profession. The Wolters Kluwer and Thomson Reuters surveys use different sampling frames (accounting broadly vs. tax specifically); percentages are not directly comparable across sources. 5. publisher_ai_policies.csv Description: Dates of first formal AI policy adoption by major academic publishers and COPE. Sources: Publisher websites, COPE position statements, and the Sify.ai AI Policies in Academic Publishing database (2025). Key columns: publisher, policy_date_decimal, policy_date_label Limitations: Policy dates are approximate (month of first issuance). Policies may have been updated since. Detailed provisions are summarised in the report; refer to original policy documents for exact wording. 6. ai_regulation_timeline.csv Description: Major global AI regulatory milestones from 2021 to 2026. Sources: European Commission, White House, NIST, Council of Europe, and Chinese government publications. Cross-referenced against the OECD AI Policy Observatory and academic reviews of AI governance. Key columns: date, event, jurisdiction Limitations: Focuses on EU, US, China, and international agreements. Does not include all national-level regulations. Dates for future milestones (2026) are per legislated timelines and may change. 7. ai_accounting_publications.csv Description: Estimated annual counts of AI-related publications in top-tier ABS 4/4* and ABDC-A* accounting and finance journals (2017-2026). Sources: Scopus and Web of Science keyword searches by the author. List of keywords and journals searched is documented in the CSV header. Key columns: year, methodological_empirical, conceptual_review, total Limitations: Keyword-based classification may misclassify papers. 2025 counts are partial due to publication lag and indexing delays. 2026 counts are projected from the 2022-2025 trend. Classification into methodological vs. conceptual is based on abstract-level screening and is subjective. 8. big4_ai_investments.csv Description: Announced AI investment commitments and platform development by the Big Four accounting firms. Sources: Firm press releases and media reports (see CSV header for specific citations: Deloitte press release 2024; PwC press release 2024, reported in Financial Times Apr 2024; EY press release Sep 2023, reported in WSJ; KPMG press release Jun 2025). Key columns: firm, ai_platform, investment_usd_billions, announcement_year, key_initiatives Limitations: Investment figures are firm-announced commitments, not independently audited spending. Commitments may span multiple years. Platform capabilities are as described in announcements and may differ from deployed functionality. ================================================================================ DATA QUALITY AND REPRODUCIBILITY ================================================================================ All datasets were compiled manually from public sources and cross-checked where possible. Data were verified against at least two independent sources for major figures (regulatory dates, adoption survey results, parameter estimates for widely-discussed models). Remaining discrepancies are noted in the limitations for each dataset. The Python replication script (scripts/replicate.py) loads these datasets and reproduces all charts and statistics in the report. For data obtained from public, open-access sources, the script reads from the CSV files (compiled versions) because the data were assembled from multiple sources and cannot be retrieved from a single URL. For proprietary model parameters and training compute, the script uses the estimated values in these CSV files and notes the estimation in chart labels. Last updated: 29 July 2026