From 9494f81e5e013fa7c790374ec8a7b3fe300d91fe Mon Sep 17 00:00:00 2001 From: ldwang Date: Thu, 7 Nov 2024 11:59:02 +0800 Subject: [PATCH] Create README.md --- examples/vl_datasets/README.md | 24 ++++++++++++++++++++++++ 1 file changed, 24 insertions(+) create mode 100644 examples/vl_datasets/README.md diff --git a/examples/vl_datasets/README.md b/examples/vl_datasets/README.md new file mode 100644 index 00000000..87bb9720 --- /dev/null +++ b/examples/vl_datasets/README.md @@ -0,0 +1,24 @@ +# Introduction + +When training the LLava-one-vision model with FlagScale, the original LLava-one-vision dataset needs to be converted to WebDataset format. This tool primarily reuses the data shuffling functionality in the original [LLava-one-vision](https://github.com/LLaVA-VL/LLaVA-NeXT/blob/main/llava/train/train.py) training trainer, while saving the data in WebDataset format on each rank. + +# Usage + +## Preparation + +1. Download https://github.com/LLaVA-VL/LLaVA-NeXT into Path_Of_LLaVA-NeXT. +2. Download google/siglip-so400m-patch14-384 into VISION_MODEL_PATH. +3. Write a hostfile with one IP per line, like the example below: +``` +1.2.3.4 slots=8 +1.2.3.5 slots=8 +``` +4. Prepare a dataset input compatible with the LLava-one-vision library, like next_ov_stage_july21.yaml. + +## Example +``` + DATA_PATH=next_ov_stage_july21.yaml + EXPNAME_PATH=*PathOfOutputWebDatasets* + HOSTFILE=hostfile + bash make_llava_ov_wds.sh $DATA_PATH $EXPNAME_PATH $HOSTFILE +```