---
license: mit
pretty_name: MoJev-Mix
task_categories:
  - multiple-choice
language:
  - en
size_categories:
  - 100K<n<1M
tags:
  - calibration
  - preference-learning
  - plackett-luce
  - structured-output
  - mojev
---

<img align="right" src="https://molemo-lab.github.io/mojev/assets/molemo-logo.png" width="76" alt="MoLeMo Lab logo">

# MoJev-Mix

[![Homepage](https://img.shields.io/badge/Project-Homepage-AD1853)](https://molemo-lab.github.io/mojev/)
[![Preprint](https://img.shields.io/badge/Preprint-PDF-AD1853)](https://github.com/MoLeMo-Lab/mojev/blob/master/paper/mojev-preprint.pdf)
[![Code](https://img.shields.io/badge/GitHub-Code-24292F?logo=github&logoColor=white)](https://github.com/MoLeMo-Lab/mojev)
[![Model](https://img.shields.io/badge/Hugging_Face-Model-FFD21E?logo=huggingface&logoColor=111111)](https://huggingface.co/MoLeMo-Lab/mojev)

Contact: [contact@molemo.org](mailto:contact@molemo.org)

The training and evaluation mixture for
[`MoLeMo-Lab/mojev`](https://huggingface.co/MoLeMo-Lab/mojev).

Each row contains state, a runtime candidate set, the correct value, and—when
available from the source domain—a graded preference over the candidates.

| MoJev family resource | location |
|---|---|
| Code | [MoLeMo-Lab/mojev](https://github.com/MoLeMo-Lab/mojev) |
| Model | [MoLeMo-Lab/mojev](https://huggingface.co/MoLeMo-Lab/mojev) |
| Dataset | **MoLeMo-Lab/mojev-mix** |
| Results | [MoJev results](https://github.com/MoLeMo-Lab/mojev#results) |
| Preprint | [MoJev (PDF)](https://github.com/MoLeMo-Lab/mojev/blob/master/paper/mojev-preprint.pdf) |
| Project page | [MoJev](https://molemo-lab.github.io/mojev/) |

## Load

```python
from datasets import load_dataset

dataset = load_dataset("MoLeMo-Lab/mojev-mix")
print(dataset["train"][0])
```

The repository also exposes the four JSONL files directly for streaming and
command-line workflows.

## Row schema

```json
{
  "context": "Choose a direction to keep the snake alive ... State: {...}",
  "options": {"answer": ["right", "down", "left"]},
  "preference": {"answer": [3, 0, 0]},
  "labels": {"answer": "right"},
  "source": "snake-v1",
  "grading": "flat"
}
```

| field | type | meaning |
|---|---|---|
| `context` | string | source state rendered as text |
| `options` | map of string arrays | candidate values for the row |
| `preference` | map of integer arrays | grades aligned with `options`; higher is better |
| `labels` | map of strings | correct value by name |
| `source` | string | source generator |
| `grading` | string | rule used to construct the grades |

`schema.json` records the field definitions and maximum candidate cardinality.

## Modalities

| modality | representation in this release |
|---|---|
| text | 507,974 rows with state serialised in `context` |
| image | state marker: `<|vision_start|><|image_pad|><|vision_end|>/absolute/path` |

The published mixture provides the text-state corpus. The same row schema accepts
multimodal decision rows: image assets are referenced from `context`, while
`options`, `preference`, and `labels` retain the same structure. Text and image
decisions therefore share one training and evaluation interface.

## Splits

| split | rows | role |
|---|---:|---|
| train | 205,084 | capped training mixture |
| validation | 82,423 | model selection and calibration |
| test | 80,703 | evaluation rows from training sources |
| `ood` | 139,764 | evaluation rows from held-out generators |

Mean candidate cardinality is 4.7. Exactly 16 candidates appear in 9.8% of
rows; 0.4% contain more than 16. A total of 38.8% of training rows carry three
or more preference levels.

## Preference signal

| grading | construction |
|---|---|
| `distance` | graph distance to the target |
| `palette` | RGB distance to the correct colour |
| `ordinal` | distance on an ordered scale |
| `priority` | source rule precedence |
| `binary`, `flat` | correct versus incorrect |

The builder converts 1,271,770 Open-Jev rows and caps each source within each
split. Capping changes painting geometry from 67.2% to 39.0% of the training
mixture and raises the share of rows with three or more grades from 28.5% to
38.8%.

## Rebuild

```sh
git clone https://github.com/MoLeMo-Lab/mojev
cd mojev
pip install -e .

mojev dataset-prepare openjev /path/to/Open-Jev/data/*/ \
  --output data/oj --max-options 64
mojev dataset-prepare balance --input data/oj --output data/mix \
  --graded-cap 60000 --plain-cap 20000
```

Both steps are deterministic:

- `mojev/openjev.py` converts source rows and constructs preference grades.
- `mojev/balance.py` applies source-level caps and writes the final splits.

## Source coverage

The mixture is derived from 18 synthetic Open-Jev generators spanning games,
navigation, document states, agent traces, and palette matching. Every row uses
the `answer` field. Palette matching supplies most graded rows. The `source`,
`grading`, and candidate cardinality fields support sliced evaluation.

The `test` split measures new rows from the training source families. The `ood`
split measures new source generators.

## Attribution

Derived from [Open-Jev](https://github.com/Zefan-Cai/Open-Jev) by Zefan-Cai et
al. under the MIT license. The grading and balancing code is also MIT licensed.