The training / finetuning code for SpatialLM is now available in the dev branch. It is adapted from LLaMA-Factory, the excellent open-source framework for LLMs finetuning. Our implementation currently includes a minimal subset of features focused on supervised fine-tuning (SFT). For additional capabilities and broader model support, please refer to the original LLaMA-Factory repository.

We also provide an example for finetuning SpatialLM1.1-Qwen-0.5B using the ARKitScenes dataset. Details can be found in FINETUNE.md. You can follow this example and adapt it to your own use cases, whether involving custom language inputs or point clouds.

Related training code request issues: #3, #5, #15, #41, #45, #57, #65