Observed arrival · 2026-10-30
A middle-schooler’s Japanese LLM experiment finds a telling mismatch
A detailed research report compares additional training methods on a Japanese language model built from scratch, with about 98.4 million parameters.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Japanese LLM researchers and model builders
- Worth noticing
- Checkpoint D had the lowest reported NLL, but a 100% loop rate across 20 generation logs and zero passes on 14 automated checks.
Field notes
The report specifies a PyTorch decoder-only Transformer with a SentencePiece BPE vocabulary of 32,768 and describes its use of public pretraining and SFT data. It separates the claim of building model weights from scratch from the source of the training data, and cautions that the short CPT run does not establish broad performance gains.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
ƒJavaScriptBrowser-side code central
One card from the complete issue