I created this dataset for Thailand’s historical minimum wage: github.com/mimocha/thailand-minimum-wage Spurred on by the recent increase in minimum wage (July 1st, 2025), I was curious about past minimum wage raises; I wanted to know how often and how much the minimum wage has changed in the past.

Then my curiosity turned into annoyance, then into sheer determination, after I repeatedly fail to find any easy to use data format. All I found were either news report citing a few previous raises, incomplete / paywalled online sources, or official government announcements which were in legalese-PDF (yuck).

Luckily, there were only 53 minimum wage change announcements (as far as I can find), and most of these are already collated into a single Excel file by the Ministry of Labor; even if it is in Thai legalese report format.

I immediately tried copying the Excel data into raw text, so I can have an LLM parse it out directly. I tried using various LLMs provided via VSCode / Github Copilot, but I quickly found this approach to be unusable. The inconsistent format of the original Excel file combined with Thai legalese reporting proved too much for even the latest Sonnet 4 and Gemini 2.5 Pro. The CSV created by these top models still hallucinated occasionally, which renders the entire dataset unusable.

However, the biggest issue isn't the hallucinations in and of itself, it is that I cannot reliably double check this CSV output manually.

The CSV file is 77+ columns by 53+ rows of pure number. Even for a dataset this small, I immediately realize how error prone my double checking attempt will be; and knew I needed another approach.

So I came up with a hybrid approach, where I manually created the CSV template first (setting up the columns and row indices), created a simple python notebook and custom python functions to fill out the CSV file row-by-row, then ask the LLM to correctly call the python functions instead.

From my experience, Sonnet 3.7 thinking & Sonnet 4 were the best at this task. After filling out the first few function calls as examples (both for myself and for the LLM), I can provide the copied text from the original Excel file as a comment across multiple notebook cells (1 announcement per notebook cell), then instruct the LLM to fill complete each cell, converting the comment into function calls.

This approach produced significantly more reliable results, and most importantly, was easily verifiable by me. I only needed to check that the function calls were correct (are the list of input province names and mapping wage value correct?), and I know the CSV file will also be correct.

All in all, this entire dataset took about 3 days of on and off working to create, with an additional day to create the Github repo and documentation.