You can retrain the instruction adapter of ChartVLM and reproduce our results via the following steps:

cd adapter/train

Training data

We have provided our training data in data.txt
You can also customize your own data with the following format:

[YOUR CUSTOMIZED SENTENCE 1] \t [LABEL 1]
[YOUR CUSTOMIZED SENTENCE 2] \t [LABEL 2]
......

Note that [Label] indicates the task type, which has been defined as:

0: Structural Extraction
1: Description/Summarization
2: Chart Type
3: Chart Title
4: Redrawing
5: Question Answering

Train

• You can train the instruction adapter using the following command:

python train_adapter.py --data_path ./path/to/data.txt --save_model_path ./path/to/model