Training a small model on my own data
AIMade by AISummarised by AI (Claude) from one piece of material it read (a talk or an explainer).How it's made
For anyone handling material that is hard to send outside, a path opens to training a judgment model yourself. One team trained Korean-language inquiry classification on an ordinary laptop and reported that accuracy, which before training was 82 of 250 questions, or 32.8%, rose to 68.8% with 100 examples in 4 minutes, 83.2% with 300 in 15 minutes, and 86.4% with 1,000 in 48 minutes. After training on 10,000 examples for 7 hours and 52 minutes, it got 598 of 600 questions right, reaching 99.7%. It also set out that one classification took 0.11 seconds, faster than going through an outside endpoint, along with an operating loop that collects the wrong cases and trains on them again. The limits are that the data is synthetic and the task is a single classification, and the figures are one team's own measurement. Another team rented hardware by the hour rather than buying it to run a large open-weight model, and put forward a break-even point: buying only pays off if you run it all day long.
Key points
- Laptop training raised classification accuracy from 32.8% to 86.4%
- After 7 hours and 52 minutes on 10,000 examples it reached 99.7% on 600 questions
- At 0.11 seconds per classification there is no need to go through an outside endpoint
- There is also the path of renting by the hour instead of buying hardware, to measure first
This is a note: a summary an AI (Claude) wrote of one piece of material it read (a talk or an explainer). It has no source link; events checked against a source are kept as records.