Official code for Language Models Meet World Models: Embodied Experiences Enhance Language Models (NeurIPS 2023). Also check our twitter here.
Install the dependencies by
pip install -r requirements.txt
All constructed training data from embodied experiences and evaluation data are released in /data/train
and /data/eval
, respectively.
Data Statistics
- Training
Task | Size |
---|---|
Plan Generation | 1659 |
Activity Recognition | 1659 |
Counting | 1000 |
Object Path Tracking | 1000 |
- Evaluation
Task | Size |
---|---|
Plan Generation | |
- Vanilla Seen | 175 |
- Vanilla Unseen | 54 |
- Confusing Seen | 135 |
- Confusing UnSeen | 43 |
Houwork QA | 261 |
Negation Housework QA | 162 |
Activity Recognition QA | 549 |
Activity Inference QA | 262 |
Counting QA | 194 |
Object Path Tracking | 200 |
Object Location QA | 200 |
We compute the fiser matrixs on the sampled 20000 examples from Pile validation set. You can download fisher-matrix-1.3B and fisher-matrix-6B from huggingface model hub, and put them under fisher-matrix
directory.
Then go to the scripts
directory where you can find all the training and evaluation scripts:
cd scripts
We provide usage examples of GPT-J-6B below. If you want to use GPJ-Neo-1.3B, just replace 6B
in the script name with 1.3B
.
If you want to train GPT-J-6B, use:
sh run_6B.sh
This script trains GPT-J-6B on a single GPU.
If you want to do distributed training, first run this:
accelerate config --config_file accelerate_config.json
and follow the instructions to set up the config file. (We also provide a sample config file in scripts
)
Then, you can simply run:
sh run_6B_multi_gpu.sh
To do evaluation on QA tasks and generation tasks, run
sh eval_qa_6B.sh
and
sh eval_gen_6B.sh
The results will be stored in output/ewc-lora-6B/qa-metric.txt
and output/ewc-lora-6B/gen-metric.txt
, respectively.
If you want to do distributed evaluation for generation tasks, please modify eval_gen_6B.sh
as:
- Replace
python eval_gen.py
withaccelerate launch --config_file accelerate_config.json eval_gen.py
- Remove
export CUDA_VISIBLE_DEVICES=0
This is same as the change from run_6B.sh
to run_6B_multi_gpu.sh
.
@article{xiang2023language,
title={Language Models Meet World Models: Embodied Experiences Enhance Language Models},
author={Xiang, Jiannan and Tao, Tianhua and Gu, Yi and Shu, Tianmin and Wang, Zirui and Yang, Zichao and Hu, Zhiting},
journal={arXiv preprint arXiv:2305.10626},
year={2023}
}