Training with SceneInstruct

We currently support LoRA fine-tuning. Please follow the instructions below to fine-tuning models with SceneInstruct as the backbone of SceneGenAgent.

Data Preparation

Download the data from Tsinghua Cloud. The data covers three tasks in SceneGenAgent: assign placement, check positional error, and fix positional error.

Fine-tuning

Take assign_placement as an example:

Preparation

1. Change the JSONL data paths of get_custom_dataset method in assign_placement_dataset.py to your own data paths
2. Change the MODEL_PATH argument in run_finetune_assign_placement.sh to your model path. You may also adjust the training hyper-parameters in this shell script.
3. If necessary, re-implement how labels_tokens is obtained in the tokenize_dialog method of assign_placement_dataset.py. This is to make sure the loss is computed only using the output in the final rounds of the conversational data. The current implementation is for the Llama-3 series, but may not suit other models.

Run training

To train the model, run:

bash run_finetune_assign_placement.sh