A project fine-tuning Google's lightweight open-weight model "Gemma 4 E2B" on a single GPU to translate Classical Korean into modern Korean has been released. Its character-level similarity score reportedly jumped from a baseline of about 4.85% to 79.93% (85.71% in one example).
Open Models · Cultural Heritage AI
Teaching a Tiny Model to Read 500-Year-Old Korean — on One GPU
Fine-tuning Google's lightweight Gemma 4 E2B to translate Classical Korean into modern Korean lifted character-level similarity from near-zero to about 80% — all on a single 16GB NVIDIA T4.
Translation accuracy — before vs. after fine-tuning
BASELINE
4.85%
character similarity
→
AFTER FINE-TUNING
79.93%
up to 85.71% in one example
A roughly 16× jump in character-level similarity after LoRA-based domain adaptation.
2.3B
effective params (≈5.1B total), 128K context
16GB
single NVIDIA T4 GPU on Colab
5
training epochs, LR 2e-5
Why it's hard
Classical Korean has no spaces between words and relies on obsolete characters and archaic grammar — Arae-a (ㆍ), Yeorin-hieut (ㆆ) — making it tough even for modern native speakers to decode.
The method
PEFT via LoRA (r=16, α=16, dropout 0.05, all-linear) makes fine-tuning feasible on 8GB–T4-class GPUs.
Training pipeline
HongGildongJeon
classical + modern pairs
→
Conversational format
prompt loss masked (-100)
→
LoRA fine-tune
Gemma 4 E2B (IT)
Promising
Korean speakers rate the classical-novel translation as practical; "the before-and-after difference is unbelievable." Unsloth reports ~1.5× faster training and 50–60% less VRAM.
Open questions
Trained on a single novel, so generalization is unverified. Tokenizer handling of spaces and obsolete characters remains a concern, and some stack issues (gibberish with use_cache=False) are reported.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…