# Training Perception Language Model (PLM) [![Hugging Face Collection](https://img.shields.io/badge/%F0%9F%A4%97%20PLM Synthetic-Image-blue)](https://huggingface.co/datasets/facebook/PLM-Image-Auto) [![Hugging Face Collection](https://img.shields.io/badge/%F0%9F%A4%97%20PLM Synthetic-Video-blue)](https://huggingface.co/datasets/facebook/PLM-Video-Auto) [![Hugging Face Collection](https://img.shields.io/badge/%F0%9F%A4%97%20PLM Human-Video-blue)](https://huggingface.co/datasets/facebook/PLM-Video-Human) We provide instruction to train or finetune PLM on a custom dataset. --- > [!TIP] > We provide configurations to run [`warm-up`](../configs/warmup/) and [`sft`](../configs/sft/) to facilitate reproducibility of PLM training. ## Data Format :open_file_folder: We use support both image and video conversation datasets using `jsonl`. Each line of `jsonl` file should follow the following format, ### For Image Conversation Dataset ```json { "image": "", "conversations": [ { "from": "human", "value": "human instruction" }, { "from": "assistant", "value": "model response" } ] } ``` ### For Video Conversation Dataset ```json { "video": "