New model weights. Exact match on the validation set goes from 90.95% → 95.3%, and held-out test / test_ood are measured and published for the first time.
| Split | n | Exact match | JSON-valid | Schema-valid |
|---|---|---|---|---|
val |
4,641 | 95.3% | 99.1% | 99.1% |
test (held out) |
4,641 | 95.2% | 99.0% | 99.0% |
test_ood |
4,775 | 94.2% | 98.8% | 98.8% |
Retrained on a 2× larger corpus (83,544 train rows, up from 42,558) on a Colab GPU. Latency on 1-thread CPU: p50 532 ms, p95 983 ms per query. Still 15.57M params — the model stays tiny and fully offline.
Upgrading
git pull && ./install.shinstall.sh now pins MODEL_VERSION=v0.2.0, so a fresh install pulls these weights automatically and verifies the sha256.
Assets
logx-model-v0.2.0.zip(58 MB) — safetensors + JSON only, no pickle fileslogx-model-v0.2.0.zip.sha256— verified automatically by the installer
Notes
- The tokenizer now adds 50 tokens (the 6 ASCII ones for JSON braces, plus accented-Latin and CJK characters present in the corpus). As always, load the tokenizer shipped in the zip, never the base model's.
- Largest remaining error buckets across all splits: top-K value errors, action confusion, and invalid JSON. Training data is synthetic —
test_oodat 94.2% is the honest generalization signal.
Full model card: MODEL_CARD.md