datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
datikz-v3
Dataset Card for DaTikZv3
DaTikZv3 is a dataset of TikZ drawings aligned with captions. In compliance with licensing agreements, certain TikZ drawings are excluded from this public version of the dataset. Check out the DeTikZify project and the DaTikZ repository for more information as well as tools and methods to recreate the complete dataset from scratch.
Usage
from datasets import load_dataset
# full dataset
ds = load_dataset("nllg/datikz-v3")
# only the train… See the full description on the dataset page: https://huggingface.co/datasets/nllg/datikz-v3.DaTikZ-V4
Dataset Card for DaTikZ-V4
DaTikZ-V4 is the dataset used to train TikZilla-3B, TikZilla-3B-RL, TikZilla-8B, and TikZilla-8B-RL for generating TikZ/LaTeX figures from natural language descriptions.
The TikZ code has been sourced from ArXiv, GitHub, and TeXStackExchange. Scientific figure descriptions were generated using Qwen2.5-VL-7B-Instruct.
Dataset fields
Each sample contains:
file_id: unique identifier
caption: original caption
vlm_description: detailed visual… See the full description on the dataset page: https://huggingface.co/datasets/nllg/DaTikZ-V4.DaTikZ-V4
Dataset Card for DaTikZ-V4
DaTikZ-V4 is the dataset used to train TikZilla-3B, TikZilla-3B-RL, TikZilla-8B, and TikZilla-8B-RL for generating TikZ/LaTeX figures from natural language descriptions.
The TikZ code has been sourced from ArXiv, GitHub, and TeXStackExchange. Scientific figure descriptions were generated using Qwen2.5-VL-7B-Instruct.
Dataset fields
Each sample contains:
file_id: unique identifier
caption: original caption
vlm_description: detailed visual… See the full description on the dataset page: https://huggingface.co/datasets/YangruiWANG/DaTikZ-V4.datikz
Dataset Card for DaTikZ
DaTikZ is a dataset of TikZ drawings aligned with captions. In compliance with licensing agreements, certain TikZ drawings are excluded from this public version of the dataset. Check out the AutomaTikZ project and the DaTikZ repository for more information as well as tools and methods to recreate the complete dataset from scratch.
Usage
from datasets import load_dataset
# full dataset
ds = load_dataset("nllg/datikz")
# only the train split… See the full description on the dataset page: https://huggingface.co/datasets/nllg/datikz.datikz-v2
Dataset Card for DaTikZv2
DaTikZv2 is a dataset of TikZ drawings aligned with captions. In compliance with licensing agreements, certain TikZ drawings are excluded from this public version of the dataset. Check out the DeTikZify project and the DaTikZ repository for more information as well as tools and methods to recreate the complete dataset from scratch.
Usage
from datasets import load_dataset
# full dataset
ds = load_dataset("nllg/datikz-v2")
# only the train… See the full description on the dataset page: https://huggingface.co/datasets/nllg/datikz-v2.datikz
Dataset Card for "datikz"
More Information needed
diagram_desc_datikz
diagram_desc_datikz — Kimi-K3 descriptions + reasoning traces for DaTikZ-V4
199,244 TikZ programs from DaTikZ-V4, each paired with
a Kimi-K3 natural-language description and a reasoning trace. Built for training text → TikZ models:
the description is the input, the TikZ code is the target. Includes the 984-item DaTikZ v2/v3 official test
split (flagged — see Test-split hygiene).
Columns
column
what it is
file_id
upstream identifier. Several naming… See the full description on the dataset page: https://huggingface.co/datasets/explcre/diagram_desc_datikz.datikz-v2-moondream-labelsdatikz_v4_rl_hybrid_428kdatikz-filtered_512_1024mm_datikzdatikz-v3-recaptioned
DaTikZ-v3 Recaptioned
A 10,000-sample subset of DaTikZ-v3 with captions regenerated using Gemini Flash via vision-language captioning.
Captioning
Original captions were replaced by passing each rendered diagram image to Gemini Flash with the prompt: describe the diagram concisely, focusing on geometric shapes, mathematical concepts, key visual elements, and purpose (~1–3 sentences starting with "A diagram ..."). Only samples with TikZ code shorter than 1000 characters… See the full description on the dataset page: https://huggingface.co/datasets/catslashbin/datikz-v3-recaptioned.datikz_test_paired_desc
datikz_test_paired_desc — the DaTikZ v2/v3 official test split, with paired descriptions and reference images
The 984-item held-out test split used to evaluate text → TikZ models, with two descriptions of every
diagram so that description style can be isolated from model quality, plus the reference image for
image-space metrics.
This is deliberately a separate repo from the training corpus
(explcre/diagram_desc_datikz), which
contains no test items. Train and test cannot be… See the full description on the dataset page: https://huggingface.co/datasets/explcre/datikz_test_paired_desc.datikz_cleaned
datikz_cleaned
The datikz__x family of the ElliotVL supervised-fine-tuning pool, after VLM cleaning.
images
6,718
QA turns
26,045
answers rewritten by the cleaning pass
303
QA created by the cleaning pass (new_qa)
22,011 (84.5%)
shards
1
How this was cleaned
A vision-language model read each image together with its QA and judged the item. The pass is
not a filter that only removes rows — it rewrites answers it finds wrong but salvageable… See the full description on the dataset page: https://huggingface.co/datasets/elliot-mllm/datikz_cleaned.datikz_merged_cleaned
datikz_merged_cleaned
The datikz_merged family of the ElliotVL supervised-fine-tuning pool, after VLM cleaning.
images
171,191
QA turns
758,950
answers rewritten by the cleaning pass
110,787
QA created by the cleaning pass (new_qa)
530,139 (69.9%)
shards
6
How this was cleaned
A vision-language model read each image together with its QA and judged the item. The pass is
not a filter that only removes rows — it rewrites answers it finds… See the full description on the dataset page: https://huggingface.co/datasets/elliot-mllm/datikz_merged_cleaned.datikz-v2datikz-filtered_0_512datikz-v2v3-test-vlm-descriptions
Reconstructed vlm_description for the DaTikZ v2/v3 official test split
Text→TikZ models in the TikZilla family are trained on DaTikZ-V4's vlm_description field — a
VLM-written visual description of the gold figure. No public split ships that field for held-out
evaluation, which makes matched-input evaluation of those models impossible off-the-shelf:
dataset
split
has vlm_description?
nllg/DaTikZ-V4
train only (no test split)
✅ yes
nllg/datikz-v3
train / test (542)
❌… See the full description on the dataset page: https://huggingface.co/datasets/explcre/datikz-v2v3-test-vlm-descriptions.DaTikZ-V4-quality-scoreddatikz-filtered_1024_2048Translated_Datikz-v3datikz_translate_testdatikz-filtered_2048_4096datikz-filtered_0_4096datikz-filtered_0_2048datikz-v2-test