---
license: cc-by-nc-4.0
library_name: transformers
pipeline_tag: text-classification
gated: manual
extra_gated_prompt: "Access to ProKope-421M weights is provided for academic evaluation, benchmarking, and commercial licensing inquiries. Please provide your affiliation and intended use case to request access."
extra_gated_fields:
  Company or Institution: text
  Corporate or Academic Email: text
  Intended Evaluation Use Case: text
language:
- en
datasets:
- LocalLLaMA/typed-decisions
tags:
- text-classification
- typed-decisions
- modernbert
- prokope
model_name: ProKope-421M
author: Alan Canto
creator: Alan Canto
---

<p align="center">
  <img src="banner.jpg" alt="ProKope-421M" width="100%" style="border-radius: 8px; margin-bottom: 16px;">
</p>

# ProKope-421M

## Introduction

**ProKope-421M** is a 421M-parameter non-autoregressive classification and scoring model developed by [ProKope AI](https://huggingface.co/ProKope-AI), based on `ModernBERT-large` with a 1,024-token context window. It evaluates structured state inputs in a single forward pass across three typed decision primitives:
1. **`noul`**: Binary boolean classification (`[0, 1]` probability).
2. **`choice`**: Multi-class categorical routing.
3. **`score`**: Ordinal severity and confidence estimation (`[0, 1]` scalar).

For 8,192-token context workloads, see [ProKope-421M-8K](https://huggingface.co/ProKope-AI/ProKope-421M-8K).

---

## Evaluation

Evaluated across 400 test cases (2,000 total decisions) on the [`LocalLLaMA/typed-decisions`](https://huggingface.co/datasets/LocalLLaMA/typed-decisions) benchmark:

| Model | Parameters | Context | Overall Accuracy | Decisions Correct | Brier Score |
| :--- | :---: | :---: | :---: | :---: | :---: |
| **ProKope-421M-8K** | 421M | 8,192 | **71.85%** | 1,437 / 2,000 | 0.384 (raw) |
| Featherless Simple Jev | 35B | 8,192 | 71.60% | 1,432 / 2,000 | 0.110 |
| **ProKope-421M** | 421M | 1,024 | 71.40% | 1,428 / 2,000 | **0.071** (calibrated) |
| prima-ratio | 12B | 8,192 | 70.20% | 1,404 / 2,000 | 0.125 |
| mgoeckel/oscar-1-400m | 400M | 1,024 | 70.00% | 1,400 / 2,000 | 0.062 |
| OpenDecider-small | 4B | 8,192 | 67.10% | 1,342 / 2,000 | 0.117 |
| Bekko System One v0 | 400M | 1,024 | 66.80% | 1,336 / 2,000 | 0.113 |

### Domain Breakdown (ProKope-421M: 71.40% / 1,428 / 2,000)
- **Invoice Processing**: 79.00% (395 / 500)
- **Agent-Trace Observability**: 73.20% (366 / 500)
- **Security Incidents**: 69.20% (346 / 500)
- **Customer Service Routing**: 64.20% (321 / 500)

### Decision Primitive Breakdown
- **`noul` (Binary Decision)**: 79.67%
- **`choice` (Multi-Class Routing)**: 69.67%
- **`score` (Ordinal Severity)**: 66.50%

---

## Model Details

| Property | Value |
| :--- | :--- |
| **Backbone** | `ModernBERT-large` (28 bidirectional encoder layers, 1,024 hidden size) + 2-layer multi-head decision projection |
| **Total Parameters** | 421,293,830 (`803.57 MB` in `bfloat16`) |
| **Context Window** | 1,024 tokens (`head_max_len`: 256 tokens) |
| **Throughput** | ~23.5 decisions/sec (~42 ms/decision on single GPU) |
| **Recommended VRAM** | 2 GB+ for `bfloat16` / `float16` inference |
| **Precision** | `torch.bfloat16` |
| **License** | `CC-BY-NC-4.0` (Commercial licensing available via ProKope AI) |

---

## Quickstart

### 1. Python Inference (`RLAgent`)

```python
from rl_agent_api import RLAgent

agent = RLAgent("ProKope-AI/ProKope-421M")

state = (
    "Production API Gateway alert: 504 Gateway Timeout spiked to 14.8% over the last 3 minutes. "
    "Worker node 4 pod eviction reported due to memory pressure (96.2%)."
)
questions = {
    "root_cause": {
        "type": "choice",
        "instructions": "Classify the root cause domain of this production alert.",
        "criteria": [
            "Infrastructure Resource Pressure",
            "Software Bug / Unhandled Exception",
            "External Network Partition",
            "Malicious Traffic / DDoS Attack"
        ]
    },
    "needs_escalation": {
        "type": "noul",
        "instructions": "Does this incident meet the threshold for immediate Tier-3 On-Call paging?",
        "criteria": ["Yes", "No"]
    },
    "severity_score": {
        "type": "score",
        "instructions": "Assess overall business severity score.",
        "criteria": ["Low", "Medium", "High", "Critical"]
    }
}

results = agent.system_one(state, questions)
for key, ans in results["answers"].items():
    print(key, "->", ans)
```

### 2. Reproducing the Benchmark Evaluation

```bash
python eval_prokope.py
```

---

## Limitations & Intended Use

- **Intended Use**: Low-latency structured classification, incident triage, workflow routing, and calibrated probability estimation over bounded decision schemas (`noul`, `choice`, `score`).
- **Limitations**: `ProKope-421M` is a non-autoregressive encoder-classifier and does not generate free-form text. Inputs exceeding the 1,024-token context window are truncated.

---

## Contact & Citation

- **Organization**: [ProKope AI](https://huggingface.co/ProKope-AI)
- **Authors**: Alan Canto, Wesley Foreman
- **Inquiries & Commercial Licensing**: [prokope.model@gmail.com](mailto:prokope.model@gmail.com)

```bibtex
@misc{canto2026prokope421m,
  author       = {Alan Canto and Wesley Foreman},
  title        = {ProKope-421M: Non-Autoregressive Structured Typed Decision Model},
  howpublished = {Hugging Face},
  year         = {2026},
  url          = {https://huggingface.co/ProKope-AI/ProKope-421M}
}
```