---
license: apache-2.0
language:
- en
- tr
- de
base_model: internlm/Intern-Decision-4B
library_name: transformers
pipeline_tag: text-classification
datasets:
- mertkayacs/jevalt-data
tags:
- decision-model
- calibration
- conformal-prediction
- uncertainty
- reasoning
- routing
- triage
- jev
- typesafe
- qwen3.5
- english
- small-language-model
- local-llm
- on-device
- text-classification
- english-llm
widget:
- text: '{"state":"Hi, I was charged twice for my March subscription: two payments
    of €29 on 3 March. Please refund the duplicate today, otherwise I will cancel.\nThanks,
    Daniel","questions":{"decision":{"type":"choice","instructions":"Which team should
    handle this ticket?","criteria":{"Billing":"payments, invoices, refunds","Technical
    support":"bugs, errors, outages","Sales":"prices, upgrades, new contracts","Account":"login,
    password, profile changes"}}},"reasoning":"off","abstain":false}'
  example_title: 'Recorded full-precision Deem-4B: support ticket, 1 October 2026'
  output:
  - label: Billing
    score: 0.947664
  - label: Technical support
    score: 0.030092
  - label: Sales
    score: 0.013086
  - label: Account
    score: 0.009158
base_model_relation: finetune
---

# Deem-4B: small English language model for local decisions

Deem-4B is one of the best open English decision models at 4B parameters: a small English language model for text classification, routing and decisions with probabilities. **Deem-4B: 94.7% accuracy; Kev-4B: 84.7%** on held-out English decisions. The test set comes from the same data pipeline as the training data.

[Try a decision in the browser](https://huggingface.co/spaces/mertkayacs/JevAlt), or run the GGUF build locally below.

## Run locally

The Q4_K_M build used **3.04 GB RAM** at a 4k context. The JevAlt server uses llama.cpp to read decision probabilities and load the matching calibration.

```sh
pip install "jevalt[serve,gguf] @ git+https://github.com/mertkayacs/jevalt" && jevalt serve --model mertkayacs/Deem-4B-GGUF --file Deem-4B-Q4_K_M.gguf
```

Send a situation, question and options to the local server:

```python
import requests

request = {
    "state": "I was charged twice for my subscription. Please refund the duplicate.",
    "questions": {
        "team": {
            "type": "choice",
            "instructions": "Which team should handle this ticket?",
            "criteria": {"billing": "payments, refunds", "technical": "bugs, outages"},
        }
    },
    "reasoning": "off",
    "abstain": False,
}
response = requests.post("http://127.0.0.1:8000/v1/systemone", json=request)
response.raise_for_status()
print(response.json())
```

Write the situation, question and options in English. The API also supports yes/no questions, ordered scores and an `unknown` option with `abstain: true`. The card widget shows a recorded answer from the full-precision model.

<details>
<summary>Use the full-precision weights with Transformers</summary>

This repository holds the full-precision weights. The JevAlt Transformers backend preserves the decision-token format and calibration:

```sh
pip install "jevalt[hf] @ git+https://github.com/mertkayacs/jevalt" && jevalt serve --backend hf --model mertkayacs/Deem-4B
```

See the [local run guide](https://github.com/mertkayacs/jevalt/blob/main/docs/local.md) for memory and client settings.

</details>

## Results and limits

On the same held-out English decisions, Deem-4B's probability error (Brier score, lower is better) is **0.091**, against **0.256 for Kev-4B**. A separate live test on 4 October 2026 scored **122/130** for Deem-4B. [Comparison files](https://huggingface.co/datasets/mertkayacs/jevalt-bench/tree/main/results/comparison) and [live requests, including misses](https://huggingface.co/datasets/mertkayacs/jevalt-bench/tree/main/results/tested) record both tests.

The models were fine-tuned with LoRA from [Intern-Decision-4B](https://huggingface.co/internlm/Intern-Decision-4B), using [JevAlt training data](https://huggingface.co/datasets/mertkayacs/jevalt-data). English is this model's focus.

- Refit calibration on your own data before using confidence thresholds. See [JevOss](https://github.com/mertkayacs/jevoss).
- The model accepts text, with an 8k context. Long irrelevant text and date arithmetic remain weak spots.
- Planted instructions still change some answers. Keep authorization checks outside the model and review consequential decisions.

## Downloads and project

[GGUF files and measured agreement](https://huggingface.co/mertkayacs/Deem-4B-GGUF) | [Code and training](https://github.com/mertkayacs/jevalt) | [Deem-4B project page](https://jevalt.mertkayacs.com/models/deem-4b/) | [Training and evaluation details](https://github.com/mertkayacs/jevalt/blob/main/docs/how-it-was-built.md).

Apache-2.0. Cite the [JevAlt repository](https://github.com/mertkayacs/jevalt) for this release.

An [Eschatia Labs](https://eschatialabs.com) project. [Mert Kaya](https://mertkayacs.com).