---
license: other
pretty_name: Bpl Cards
task_categories:
  - image-to-text
tags:
  - paratext
  - glam-eval
  - library-metadata
  - bpl-cards
size_categories:
  - n<1K
---
# Bpl Cards

Catalogue metadata extracted from digitised material with a multimodal model and
human-reviewed.

## Schema

| Field | Type | Description |
| --- | --- | --- |
| `call_number` | `string` | |
| `heading` | `string` | |
| `author` | `string` | |
| `title` | `string` | |
| `imprint` | `string` | |
| `collation` | `string` | |
| `notes` | `string` | |
| `tracings` | `string` | |

The exact machine-readable output schema (enum values, nested objects, optionality)
is in [`schema.json`](schema.json) — the `response_format` the labels follow.

## How it was labelled

- **Model:** `Qwen/Qwen3-VL-30B-A3B-Instruct`
- **Prompt hash:** `6993b107dd8d`
- **Schema version:** `v1`
- **Review:** each item was shown to a human reviewer who gave a verdict and an
  optional free-text note. **Gold labels are the 2 items a human signed
  off:** 2 _verified_ (model output a reviewer marked good enough) and
  0 _corrected_ (a reviewer edited the fields into the right answer).
  Each row's `_label_status` (`verified` / `corrected`) and, for corrected rows,
  `_corrected_fields` record which is which.
- **Model baseline (for benchmarking — not the label quality):** the labels above are
  fully human-verified, so this figure describes the *generating* model, not the
  dataset. It agreed with reviewers on **70.0%** of 5
  scored items (good_enough=2,
  needs_tweaks=3,
  not_accurate=0; needs_tweaks scored as half credit) —
  a reference point for scoring other models against this set.

<details>
<summary>Extraction prompt</summary>

```
You are reading a catalogue card from the Boston Public Library's Rare Books
Department. Look at the image and return a single JSON object matching the
schema. Emit JSON only — no commentary or markdown.

- `call_number`: the shelfmark, usually printed down the left-hand edge.
- `heading`: the subject or added heading, usually typed in capitals across the
  top of the card. Leave it null when the card has none.
- `author`: the main entry — the personal or corporate name the card files
  under, transcribed as printed, including dates.
- `title`: the title of the work.
- `imprint`: place, publisher and date of publication.
- `collation`: the physical description — pagination, illustrations, size.
- `notes`: anything else the card records, such as a contents list.

- `tracings`: the numbered list of added entries at the foot of the card.
  Transcribe it as printed. Null if the card has none.

Transcribe as printed and never add markup. Use null for anything the card does
not carry — never infer a `heading` from the title.
```

</details>

## Provenance

Produced by paratext. Each row carries `_label_status`, `_verdict`,
`_prompt_hash`, `_schema_version`, `_round`, `_model`, and the reviewer's
`_review_note` (plus `_corrected_fields` on corrected rows).

## Rights & license

No licence set (front matter records `license: other`). **CC0-1.0** is recommended for open sharing — pass `--license`/`--rights` to `paratext export` or set them under `[project.<name>.export]`.

## Limitations

Single-institution scope; labels reflect one model + prompt at a point in time and
the reviewing institution's cataloguing conventions.