CYBRTRKK/certo-v1

Certo v1

Four 421M-parameter checkpoints for typed decisions, tool routing, web computer use, and breadth classification. Each is a ModernBERT-large encoder with a shared option scorer over (state, question, option) items; the checkpoints differ only in what they were last trained on.

Checkpoints

folder	skill	held-out result
typed-decisions/	typed decisions, 2,000 decisions	0.7715 overall; choice 0.7383, score 0.7475, noul 0.8367
routing/	tool routing, 5,575 cases	0.8872 route; 0.9883 needs_tool
computer-use/	Mind2Web Cross-Domain, 3,911 steps	0.6863 element; 0.9665 operation; 0.8264 overall
breadth/	four-epoch breadth fine-tune for the classification rows	see questions.json for the question bank

The first three numbers are re-read from each run's score_*.json / td_*.meta.json; the full write-up, including the external comparisons and their caveats, is reports/MODEL_RELEASE.md in the training repository.

Files

Every folder contains the same layout:

• model.safetensors — full checkpoint, 421.3M parameters, F16 (804 MB)
• rl_agent_config.json — encoder, sequence budgets, fitted temperatures, abstention thresholds
• tokenizer/ — fast tokenizer
• encoder/config.json — ModernBERT-large configuration
• questions.json — the question bank the checkpoint was fitted against
• train_report.json — the run's own in-sample calibration report

Loading

# laya 0.4.2 (Apache-2.0), the package these checkpoints were trained with
from laya.train import load_checkpoint

model, tokenizer, cfg = load_checkpoint("typed-decisions")

The base encoder is answerdotai/ModernBERT-large (Apache-2.0); its configuration is included in each folder. laya pulls its own torch, transformers and huggingface_hub dependencies.

What these are not

• Computer use is candidate reranking. Mind2Web supplies eight candidates and the checkpoint ranks them; a deployed agent still needs a recall stage.
• Routing's edge is the decline option. On rows where a tool is gold the checkpoint is level with a zero-shot 421M specialist; the gain is in no_tool.
• Fitted, in-distribution numbers. The external rows in the release note are zero-shot and are not like-for-like wins.
• No destructive-action or confirmation label was trained. Mind2Web has no such label; that is separate work.

License

The weight files in this repository are released under the Apache License 2.0.

Copyright 2026 the Certo authors.

The base encoder is answerdotai/ModernBERT-large (Apache-2.0); its configuration is included in each folder. Training sources are xLAM-60k, xLAM-irrelevance, ToolACE, Seal-Tools, Mind2Web and this project's gold corpus. Their licenses are permissive: Apache-2.0, MIT and CC-BY-4.0.

If you redistribute these weights, keep the license and copyright notice with them.