CrochetBench Dataset

CrochetBench: Can Vision-Language Models Move from Describing to Doing in Crochet Domain? ACL 2026 (Long Paper)

This dataset is the benchmark split used in the CrochetBench paper to evaluate vision-language models (VLMs) on crochet pattern understanding and generation tasks.

• Paper: ACL Anthology (link to be added after conference)
• Code: github.com/Peiyu-Georgia-Li/crochetBench
• Fine-tuned model (Task C): peiyugeorgia/qwen2vl-crochetbench-task-c-finetuned

---

⚖️ License

This dataset is licensed under CC BY-NC 4.0.

The crochet patterns in this dataset are derived from content published by Yarnspirations (brands: Bernat, Patons, Lily Sugar'n Cream, Caron, Red Heart, Sugar Bush, Phentex).

Yarnspirations has granted permission for use of their patterns in academic research and personal use. The following restrictions apply:

• ✅ Academic research and non-commercial study
• ✅ Personal use
• ✅ Share and adapt with attribution and a link to this license
• ❌ Reproduction or resale for commercial purposes
• ❌ Use in commercial products or services

Attribution: If you use this dataset, please cite the CrochetBench paper (see Citation) and acknowledge Yarnspirations as the original content source.

Note: The image_link field contains URLs to product images hosted by Yarnspirations at yarnspirations.com. Image access is subject to Yarnspirations' own terms of service.

---

📊 Dataset Summary

Split	File	Examples
Train	train.json	4,859
Validation	val.json	274
Test	test.json	877

Total: 6,010 crochet pattern examples sourced from multiple Yarnspirations brands.

Top project categories (train): Afghans/Blankets, Hats, Sweaters/Cardigans, Shawls/Wraps, Scarves.

---

🧶 Tasks

CrochetBench evaluates VLMs on four tasks using crochet pattern images:

Task	Name	Description
A	Stitch Recognition	Identify crochet stitches present in an image
B	Instruction Selection	Select the correct pattern instructions for an image
C	Instruction Generation	Generate crochet instructions from an image
D	Instruction-to-DSL	Convert crochet instructions to the CrochetPARADE DSL

---

📁 Data Fields

train.json and val.json

Field	Type	Description
id	string	Yarnspirations pattern ID (e.g. RHC0502-014740M)
pattern_name	string	Title of the crochet pattern
project_type	string	Category (e.g. Afghans_Blankets, Hats, Scarves)
skill_type	string	Always "crochet"
skill_level	string	Skill level (e.g. beginner, intermediate, experienced)
materials	list[string]	Yarn, hook, and notions
abbreviations	list[string]	Stitch abbreviations used in the pattern
measurements	string	Finished project dimensions
gauge	string	Gauge / tension information
instructions	string	Full crochet pattern instructions
image_link	string	URL to the Yarnspirations product image
source	string	Original source PDF path (local reference, not portable)
source_file	string	Category JSON the entry was sourced from

test.json (additional fields)

Field	Type	Description
DSL	string	CrochetPARADE DSL representation (where available)
prompt	string	Task prompt used for evaluation
generated_instructions	string	Pre-generated instructions for Task D evaluation
source_category	string	Category label

---

🔧 Usage

import json

with open("train.json") as f:
    train = json.load(f)

example = train[0]
print(example["pattern_name"])   # e.g. "Cozy Heart Afghan"
print(example["image_link"])     # Yarnspirations image URL
print(example["instructions"][:200])

Or load with the datasets library:

from datasets import load_dataset

ds = load_dataset("peiyugeorgia/CrochetBench")
print(ds["train"][0])

---

📝 Citation

If you use this dataset, please cite:

@inproceedings{peiyuli2026,
  title     = {CrochetBench: Can Vision-Language Models Move from Describing to
               Doing in Crochet Domain?},
  author    = {Li, Peiyu and Huang, Xiaobao and Hua, Ting and Chawla, Nitesh V.},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for
               Computational Linguistics (Volume 1: Long Papers)},
  year      = {2026},
  note      = {To appear}
}