Model Card for ChartGPT
Model Details
Model Description
This model is used to generate charts from natural language. For more information, please refer to the paper.
- Model type: Language model
- Language(s) (NLP): English
- License: Apache 2.0
- Finetuned from model: FLAN-T5-XL
- Research paper: ChartGPT: Leveraging LLMs to Generate Charts from Abstract Natural Language
Model Input Format
Model input on the Step x. Specifically, <...> serves as a seperation token.
{table name}
<head> {column names}
<type> {column types}
<data> {data row 1} <line> {data row 2} <line>
<utterance> {NL utterance}
<ans>
<sep> {Step 1 prompt} {Answer 2}
...
<sep> {Step x-1 prompt} {Answer x-1}
<sep> {Step x prompt}
And the model should output the answer corresponding to step x.
The step 1-6 prompts are as follows:
Step 1. Select columns:
Step 2. Add filter:
Step 3. Add aggregations:
Step 4. Select chart type:
Step 5. Choose encoding:
Step 6. Add sort:
How to Get Started with the Model
Running the Model on a GPU
An example of a movie dataset with an utterance "What kinds of movies are the most popular?". The model should give the answers to step 1 (select columns). You can use the code below to test if you can run the model successfully.
from transformers import (
AutoTokenizer,
AutoModelForSeq2SeqLM,
)
tokenizer = AutoTokenizer.from_pretrained("yuan-tian/chartgpt")
model = AutoModelForSeq2SeqLM.from_pretrained("yuan-tian/chartgpt", device_map="auto")
input_text = "movies <head> Title,Worldwide_Gross,Production_Budget,Release_Year,Content_Rating,Running_Time,Major_Genre,Creative_Type,Rotten_Tomatoes_Rating,IMDB_Rating <type> nominal,quantitative,quantitative,temporal,nominal,quantitative,nominal,nominal,quantitative,quantitative <data> From Dusk Till Dawn,25728961,20000000,1996,R,107,Horror,Fantasy,63,7.1 <line> Broken Arrow,148345997,65000000,1996,R,108,Action,Contemporary Fiction,55,5.8 <line> <utterance> What kinds of movies are the most popular? <ans> <sep> Step 1. Select the columns:"
inputs = tokenizer(input_text, return_tensors="pt", padding=True).to("cuda")
outputs = model.generate(**inputs)
print(tokenizer.decode(outputs[0], skip_special_tokens = True))
Training Details
Training Data
This model is Fine-tuned from FLAN-T5-XL on the chartgpt-dataset.
Training Procedure
Plan to update the preprocessing and training procedure in the future.
Citation
BibTeX:
@article{tian2024chartgpt,
title={ChartGPT: Leveraging LLMs to Generate Charts from Abstract Natural Language},
author={Tian, Yuan and Cui, Weiwei and Deng, Dazhen and Yi, Xinjing and Yang, Yurun and Zhang, Haidong and Wu, Yingcai},
journal={IEEE Transactions on Visualization and Computer Graphics},
year={2024},
pages={1-15},
doi={10.1109/TVCG.2024.3368621}
}