README.md

---
license: mit
language:
  - sv
task_categories:
  - text-classification
tags:
  - synthetic
  - swedish
  - one-pass-scorer
  - jev
  - system-one
  - form-filling
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl.gz
      - split: validation
        path: data/validation.jsonl.gz
      - split: test
        path: data/test.jsonl.gz
---

# One-Pass SV-Forms synthetic corpus (Swedish)

**This dataset is entirely synthetic.** It was generated by a script from a concept catalogue, not
collected from people or customer submissions. Names and organisations are constructed,
email addresses use `.invalid`, and identifier-shaped values are generated locally.
They are not checked against registries; coincidental matches with real names or
identifiers cannot be ruled out. The dataset is published so the recipe behind
[`precisit/one-pass-sv-forms`](https://huggingface.co/precisit/one-pass-sv-forms) can be
reproduced and refuted, not because the data has value as data.

## What it is

Training data for a **one-pass option scorer**: a model that receives a context and a supplied list
of candidate decisions and returns one score per option, generating no text. Each line is one
decision:

```json
{"context": "UPPGIFT fyll i formuläret från dokumentet och skicka sedan in\nFORM ...\nELEMENT Edit \"E-postadress\" value=\"\"",
 "options": ["fyll E-post: anna@example.invalid", "fyll Förnamn: Anna", "kryssa", "klicka", "hoppa över"],
 "label": 0,
 "meta": {"action": "fill", "form_signature": "adress|efternamn|epost|fornamn|...", "role": "Edit", "seed": 2026}}
```

- `context`: the task line, the form title and the element in question, plus the entities
  extracted from an attached document. Bytes are what the model sees; there is no tokenizer.
- `options`: the label is an index into this list, so the option order *is* part of the sample.
- `meta.action`: `fill` / `check` / `click` / `skip`, the higher-level decision.
- `meta.form_signature`: the sorted field set of the form the row came from. **The splits are made
  on this value**, so a form's rows never straddle train/validation/test.

## How it was generated

```bash
git clone https://github.com/precisit/one-pass-specialists
cd one-pass-specialists
git checkout 6a6def7887ea13a323049a972409b19aed953559
# After installing the toolkit dependencies described in its README:
python -m onepass.synth --output data/sv --episodes 10000 --seed 2026
```

With the same generator revision and environment, generation is deterministic given `seed`; `manifest.json` records the episode count, the
ratios, the generator configuration and a SHA-256 per split, so a regeneration can be checked
rather than trusted. This upload used that recipe, and the files here are gzipped copies
of its splits. See `manifest.json` for the recorded checksums and split metadata.

| Split | Episodes | Decisions | SHA-256 |
| --- | ---: | ---: | --- |
| train | 7 984 | 234 921 | in `manifest.json` |
| validation | 1 007 | 29 649 | in `manifest.json` |
| test | 1 009 | 29 839 | in `manifest.json` |
| `demo-handwritten.jsonl` | n/a | 50 | hand-written, not generated: three forms with gold decisions, kept outside the splits as an out-of-distribution check |

Action mix in the training split: `skip` 112 006, `fill` 108 147, `click` 7 984, `check` 6 784 (this
mixture is a property of the generator, not a measured distribution of real form tasks).

## Intended use and limits

- **Use it for:** reproducing the pipeline, learning the input contract, and testing whether a
  small Swedish scorer is the right component for a bounded interface task. It is the "data leg" of
  the recipe in `precisit/one-pass-specialists`.
- **Do not use it as a benchmark of production behaviour.** It is generated from a 50-concept
  catalogue with a fixed set of form templates; real Swedish forms have vocabulary, layouts and
  edge cases it does not contain. The released model's own card states the same limit, and the
  numbers on synthetic held-out rows are not a promise about any real form.
- **Data-size comparison.** At 900 episodes the same pipeline and the same schedule
  score 77 % top-1; at 4 000, 98.94 %; at 10 000, 99.29 %. The corpus is published at 10 000
  episodes because that is what trained the released checkpoint, and the smaller configurations are
  one command away if you want the curve for yourself. Each run has its own generated
  test set, and more episodes at fixed epochs mean more optimization steps. These
  results do not isolate the effect of data size or establish a learning plateau.
- **Keep real submissions out of this repository.** The corpus is generated for research.
  Do not add personal data, customer data or real document contents.

## Licence

MIT, like the generator and the model. Please cite the toolkit repository rather than this dataset
alone if you build on it.