# Jesus Rodriguez (systemonedev): kenning-large

> A 435M DeBERTa-v3-large cross-encoder by Jesus Rodriguez for noul, choice and score questions over program state. Each option is scored against the state and returned with a temperature-calibrated probability, with no generated text. Apache-2.0 weights.

- Page: https://systemonemodels.ai/systemonedev/kenning-large
- API: https://api.systemonemodels.ai/v1/models/systemonedev/kenning-large
- Download: `pip install systemonemodels && systemone pull systemonedev/kenning-large`

## Facts

| | |
|---|---|
| Maker | Jesus Rodriguez (systemonedev) (https://systemonemodels.ai/systemonedev) |
| Decides | choice, score, noul, classify |
| Architecture | kenning |
| Base model | moritzlaurer/deberta-v3-large-zeroshot-v2.0-c |
| Parameters | 435M |
| Context | 512 tokens |
| Licence | apache-2.0 |
| Availability | Open weights |
| Released | 2026-10-06 |
| Latest version | 0.5.0 |

## Reported evaluation

Suite: SystemOne Builder general suite (1,328 held-out items, 7 families, macro accuracy), the maker's own. Numbers are the publisher's own.

- Decision accuracy: 67.9%

- Median latency: 35 ms

## Model card

<!-- generated by scripts/seed_catalog.py; edit content/models/catalog.yaml -->

# kenning-large-v0.5

A 435M DeBERTa-v3-large cross-encoder by Jesus Rodriguez for noul, choice and score questions over program state. Each option is scored against the state and returned with a temperature-calibrated probability, with no generated text. Apache-2.0 weights.

Built with the maker's SystemOne Builder toolkit and served through its /v1/systemone service or the systemone-client package; systemone.dev is the maker's own community site, not affiliated with System One Models. Each (state, option) pair is scored separately, up to 512 tokens per pair, so longer state is truncated. On the maker's general suite it scores 0.679 macro accuracy (v0.4: 0.625); short text and conversation are its strongest families, while records, tables and logs score 0.52 to 0.59. Temperatures were fitted on the training distribution, so refit before relying on the probabilities. v0.4 was distilled from Cloudflare Clef-flash soft labels. Training sources are listed in NOTICE.md, some CC-BY or CC-BY-SA, so NOTICE.md travels with the weights.

## What it decides

- **choice** — picks one option from a set
- **score** — places the input on an ordered scale
- **noul** — answers a yes/no question with one probability
- **classify** — assigns a category from a fixed taxonomy

## At a glance

| | |
|---|---|
| Parameters | 435M |
| Base model | `MoritzLaurer/deberta-v3-large-zeroshot-v2.0-c` |
| Maker | Jesus Rodriguez (systemonedev) |
| Released | 2026-10-06 |
| License | apache-2.0 |
| Reported accuracy | 67.9% |
| Reported latency | 35 ms p50 on one RTX 3090 |

## Get the weights

```bash
pip install systemonemodels
systemone pull systemonedev/kenning-large
```

The files are served from the maker's Hugging Face repository, [`systemonedev/kenning-large-v0.5`](https://huggingface.co/systemonedev/kenning-large-v0.5), and verified against the checksums recorded here.

## Read more

- [Model card](https://huggingface.co/systemonedev/kenning-large-v0.5)
- [kenning-large-v0.4](https://huggingface.co/systemonedev/kenning-large-v0.4)
- [SystemOne Builder](https://github.com/systemonedev/systemone-builder)

---

*This page was opened by System One for Jesus Rodriguez (systemonedev), who can claim the organisation and take it over at any time.*

---

From System One Models — https://systemonemodels.ai/ · every System One model: https://systemonemodels.ai/system-one-models