Skip to content

DATA.md

Latest commit

 

History

History
92 lines (69 loc) · 2.52 KB

DATA.md

File metadata and controls

92 lines (69 loc) · 2.52 KB

Dataset Guide — U-RxnDiagram-15k

Overview

U-RxnDiagram-15k (formerly RxnCaption-15k) is a dataset of ~15,000 chemical reaction diagram images extracted from scientific PDFs, each annotated with:

  • Bounding boxes for all chemical entities (structures, text, identifiers, supplements)
  • Complete reaction graphs linking reactants → conditions → products

HuggingFace: songjhPKU/U-RxnDiagram-15k

Diagram Types

Type Description
single Single linear reaction (A → B)
multiple Multiple parallel or sequential reactions
tree Tree-structured multi-step synthesis
graph Complex reaction network

JSON Schema

Both the ground-truth and prediction files use this schema:

{
  "licenses": [],
  "info": {},
  "categories": [
    {"id": 1, "name": "structure"},
    {"id": 2, "name": "text"},
    {"id": 3, "name": "identifier"},
    {"id": 4, "name": "supplement"}
  ],
  "images": [
    {
      "id": 0,
      "file_name": "10.1002_foo_figure_0.png",
      "diagram_type": "single",
      "bboxes": [
        {"id": 0, "category_id": 1, "bbox": [x, y, w, h]},
        {"id": 1, "category_id": 2, "bbox": [x, y, w, h], "text": "NaOH"},
        {"id": 2, "category_id": 1, "bbox": [x, y, w, h]}
      ],
      "reactions": [
        {"reactants": [0], "conditions": [1], "products": [2]}
      ]
    }
  ]
}

Category IDs

category_id Meaning
1 Chemical structure (molecule, complex, etc.)
2 Text label (reagent, solvent, temperature, yield, ...)
3 Compound identifier (number, letter, code)
4 Supplementary annotation (ignored in evaluation)

Bbox format

[x, y, width, height] in pixels (COCO-style, 0-indexed top-left origin).

Download

from datasets import load_dataset
ds = load_dataset("songjhPKU/U-RxnDiagram-15k")

Or with HuggingFace CLI:

huggingface-cli download songjhPKU/U-RxnDiagram-15k --repo-type dataset --local-dir data/U-RxnDiagram-15k

Evaluation Benchmarks

The paper evaluates on two held-out test sets:

Benchmark Source # Images
PDF-400 Internal PDF corpus 400
RxnScribe RxnScribe dataset varies

Data Processing Pipeline

Raw PDF → cropped reaction images → YOLO detection → BIVP annotation → Qwen training JSONL

See scripts/prepare_data.sh for the full automated pipeline.