Repository navigation
Releases
Releases · leofmarciano/groot
Groot 1 · v0.1.1 — Folha correction
Groot 1 v1.1 fixes the zero-shot failure reported after launch.
| Groot 1 (S+M) | Groot 1 S | Jev | Julia 1 | |
|---|---|---|---|---|
| Folha news pt · zero-shot, 400 | 87.8% | 84.5% | 85.8% | 49.5% |
| Typed decisions · 2,000 | 75.3% | 75.2% | 73.7% | 72.7% |
- Groot 1 averages a 149M S and a new 333M M model (mmBERT-base): marcyano/Groot-1.
- Groot 1 S weights replaced with the corrected model: marcyano/Groot-1-S. v1.0 scored 33.8% on Folha.
- GrootEngine loads any directory with an
ensemble.jsonas a probability-averaged ensemble. - New training data: label-diversified topic sets, URL-labelled Portuguese web news (Folha/UOL blocked), Qwen2.5-14B soft labels. No Folha text and no Jev outputs in training.
- The lead over Jev on Folha is 8 of 400 articles, within the ±3-point margin. Selection, test runs and remaining gaps are in the README.
Groot 1 · v0.1.0
First public release of Groot 1, an open typed-decision model by Marciano Theory SI.
Highlights
- Typed decisions (2,000 questions): Groot 1 S 75.2% · Jev 73.7% · Julia 1 72.7%
- Calibration: typed-decision Brier 0.072, against 0.156 for Jev and 0.336 for Julia 1
- Long context: 96–100% needle accuracy from 1k to 64k tokens
- System One compatible server (
/v1/systemone), up to 255 options per call - Pure Bend engine matching PyTorch to within 1e-7
What's inside
groot1/: model, serializer with BM25 state retrieval, trainer, evaluation, server, exportbend/: the full forward pass in Bend (JS host and native binary)scripts/benchmark.py: three-way benchmark harness; raw results inmetrics/
Known limits
Classification results are in-domain. Groot trails Jev on BoolQ, ASSIN2 and HateBR, and is slower than Jev's API above 16k tokens. Weights will be published separately. See the README for the full protocol.
🤖 Generated with Claude Code