Chinese BabyLM A195 Hanzi Best
Current verified Hanzi-track model for the Chinese BabyLM challenge.
- Backend:
causal
- Model type: custom log-prob ensemble wrapper over strict official-corpus checkpoints
Hanzi Scores
Open Tasks
| Task | Score |
|---|
| hanzi_structure | 0.5855 |
| hanzi_pinyin | 0.5610 |
Hidden Tasks
| Task | Score |
|---|
| hanzi_structure_hidden | 0.5830 |
| hanzi_pinyin_hidden | 0.5340 |
Competition Compliance
- Trained from random initialization
- Official Chinese BabyLM corpus only (
chinese-babylm-org/babylm-zho-100M)
- No external pretrained checkpoint
- No external linguistic resource (no pypinyin, IDS/CJKVI, radical tables, glyph/font resources, or inference-time rules)
- No public/hidden eval item text, labels, predictions, or UNK lists read
- No inference-time rules
Usage
Use with trust_remote_code=True. The repo includes all source submodels under submodules/; the wrapper resolves those paths relative to this repository snapshot.
The challenge config.yaml is included at the repository root.
