# Hemant Kumar: verdict

> A 151M ModernBERT encoder by Hemant Kumar, post-trained from Knowledgator's GLiClass on Banking77 intents with cross-entropy plus Brier loss. It scores up to 24 typed options plus an abstain slot in one pass; strong on banking, weak on general decision suites by its own numbers.

- Page: https://systemonemodels.ai/heman10x/verdict
- API: https://api.systemonemodels.ai/v1/models/heman10x/verdict
- Download: `pip install systemonemodels && systemone pull heman10x/verdict`

## Facts

| | |
|---|---|
| Maker | Hemant Kumar (https://systemonemodels.ai/heman10x) |
| Decides | choice, score, noul |
| Architecture | gliclass |
| Base model | knowledgator/gliclass-modern-base-v2.0 |
| Parameters | 151M |
| Context | 512 tokens |
| Licence | apache-2.0 |
| Availability | Open weights |
| Released | 2026-09-17 |
| Latest version | 2026.09 |

## Reported evaluation

Suite: Held-out Banking77 split (1,000 cases, 5 candidates), the maker's own. Numbers are the publisher's own.

- Decision accuracy: 95.0%
- Calibration error (ECE): 0.034
- Median latency: 35.6 ms
- p95 latency: 39.8 ms

## Model card

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

# OpenJev Verdict (rlcd-modernbert-151m)

A 151M ModernBERT encoder by Hemant Kumar, post-trained from Knowledgator's GLiClass on Banking77 intents with cross-entropy plus Brier loss. It scores up to 24 typed options plus an abstain slot in one pass; strong on banking, weak on general decision suites by its own numbers.

Not RLCD-trained, despite the repo name. Trained only on banking intents and CLINC150 out-of-scope data; by the maker's own numbers general-domain accuracy is low (48.07% on TypeSafe's public evaluation). Abstention degrades on hard negatives, option order flips 3 to 4.5% of answers, and calibration was fitted on banking plus synthetic slices only. The card and the GitHub README disagree on the uncalibrated ECE. Needs the gliclass package and the maker's rlcd SDK. ONNX fp32 and fp16 builds are in the repo. A later "Verdict 2.0" has no released 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

## At a glance

| | |
|---|---|
| Parameters | 151M |
| Base model | `knowledgator/gliclass-modern-base-v2.0` |
| Maker | Hemant Kumar |
| Released | 2026-09-17 |
| License | apache-2.0 |
| Reported accuracy | 95.0% |
| Reported latency | 35.58 ms p50 / 39.81 ms p95 (hardware not stated) |

## Get the weights

```bash
pip install systemonemodels
systemone pull heman10x/verdict
```

The files are served from the maker's Hugging Face repository, [`heman10x/rlcd-modernbert-151m`](https://huggingface.co/heman10x/rlcd-modernbert-151m), and verified against the checksums recorded here.

## Read more

- [Model card](https://huggingface.co/heman10x/rlcd-modernbert-151m)

---

*This page was opened by System One for Hemant Kumar, 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