---
license: cc-by-nc-4.0
task_categories:
- image-text-to-text
- visual-question-answering
language:
- en
tags:
- patent
- patent-examination
- multimodal
- legal
- uspto
- novelty
- non-obviousness
pretty_name: PANORAMA-NOC4PC-Multimodal
size_categories:
- 100K<n<1M
configs:
- config_name: noc4pc
  data_files:
  - split: train
    path: noc4pc/train-*
  - split: validation
    path: noc4pc/validation-*
  - split: test
    path: noc4pc/test-*
- config_name: drawings
  data_files:
  - split: train
    path: drawings/train-*
---

# PANORAMA-NOC4PC-Multimodal

A **multimodal extension** of the NOC4PC benchmark from
[PANORAMA](https://huggingface.co/datasets/LG-AI-Research/PANORAMA)
(Lim et al., *PANORAMA: A Dataset and Benchmarks Capturing Decision Trails and
Rationales in Patent Examination*, NeurIPS 2025 Datasets & Benchmarks).

The original NOC4PC task asks a model to judge a patent claim's patentability
(**§102 / §103 / allow**) from **text only** (claim + cited prior-art text).
This dataset adds the **drawings** that real examiners actually look at — for
**both the patent application and every cited prior-art reference** — so the
task can be studied as a true vision-language problem.

## Why two configs (reference layout)

In NOC4PC the same drawing is shared by many rows (a single application appears
in ~18 claim-level rows on average; popular prior-art patents appear in hundreds).
Embedding the image in every row would duplicate it dozens of times and blow the
dataset up past 600 GB. Instead — exactly like VQAv2 / DocVQA / OK-VQA — we store
**each unique drawing once** and let task rows **reference it by `image_id`**:

| config | one row = | key columns |
|---|---|---|
| `noc4pc` | one claim-level examination instance | `..., application_image_ids[], prior_art_image_ids[]` |
| `drawings` | one rendered drawing **page** | `image_id, source, ref_number, page, image` |

`image_id` format: `app_{applicationNumber}_p{N}` (application) and
`cited_{patentNumber}_p{N}` (prior art), where `N` is the 1-based page number.
Drawings are rendered from the original USPTO drawing PDFs at 120 dpi (all pages).

## Columns

**`noc4pc`**
- `application_number`, `claim_number` — the target claim
- `context` (JSON string) — application title / abstract / claims
- `prior_art_specifications` (JSON string) — cited prior art (title, abstract, claims, cited paragraph)
- `answer` (JSON string) — `{"code": "102" | "103" | "ALLOW", "reason": ...}` (gold label)
- `application_image_ids` (list[str]) — drawing pages of the application
- `prior_art_image_ids` (list[str]) — drawing pages of all cited prior arts

**`drawings`**
- `image_id` (str), `source` (`application` | `prior_art`), `ref_number` (str),
  `page` (int), `image` (`datasets.Image`)

## Usage

```python
from datasets import load_dataset

REPO = "sungjae98/PANORAMA-NOC4PC-multimodal"

# 1) load both configs
noc = load_dataset(REPO, "noc4pc", split="test")
draw = load_dataset(REPO, "drawings", split="train")   # the image bank

# 2) build an image_id -> PIL.Image lookup (lazy; decode on access)
id2idx = {iid: i for i, iid in enumerate(draw["image_id"])}
def get_image(image_id):
    return draw[id2idx[image_id]]["image"]   # PIL.Image

# 3) assemble one multimodal example
ex = noc[0]
app_imgs   = [get_image(i) for i in ex["application_image_ids"]]
prior_imgs = [get_image(i) for i in ex["prior_art_image_ids"]]
print(ex["application_number"], ex["claim_number"], "->", eval(ex["answer"])["code"])
print("application pages:", len(app_imgs), " prior-art pages:", len(prior_imgs))
# feed ex["context"] + ex["prior_art_specifications"] + app_imgs + prior_imgs to a VLM
```

> Tip: for memory-tight training, don't materialize all of `draw`; index it once and
> decode images on demand, or filter `draw` to the `image_id`s you need per batch.

## Splits (noc4pc)

| split | rows |
|---|---|
| train | 136,211 |
| validation | 7,392 |
| test | 2,884 |

## Provenance & license

Drawings and text are derived from the public USPTO record and the PANORAMA
dataset. Released under **CC-BY-NC-4.0**, following the original PANORAMA license.
This is an unofficial, research-oriented extension and is not affiliated with the
original authors or LG AI Research.

If you use this, please cite the original PANORAMA paper.