search

AILab-CVC / SEED-X

Multimodal Models in Real World
Other
392 stars 16 forks source link

About pretrain code #5

Closed HalvesChen closed 5 months ago

HalvesChen commented 5 months ago

Hi, thanks for your greate work. I would like to know if you will release pre training code？

geyuying commented 5 months ago

Hi, the pretraining code is the same as the instruction tuning execpt for:

  1. Change the pretrained model of "llm_seed_x.yaml" to llama2 pre-trained model;
  2. Change "target: src.models.mllm.seed_x.ContinuousLVLM.from_pretrained" in "agent_seed_x.yaml" to "target: src.models.mllm.seed_x.ContinuousLVLM" and remove "pretrained_model_path"
  3. Remove the instruction prompt in the data config for pre-training.
geyuying commented 5 months ago

The only differences between pre-training and instruction tuning are the pre-trained model and the data loaders.

HalvesChen commented 5 months ago

Thanks~