We introduce Text-to-text Self-conditioned Simplex Diffusion (TESS), a text diffusion model that is fully non-autoregressive, employs a new form of self-conditioning, and applies the diffusion process on the logit simplex space rather than the typical learned embedding space.

How to setup the environment

conda env create -f environment.yaml --prefix  ${LOCAL_DIR}/conda/envs/sdlm
python setup develop

to update environment after installation:

conda env update --file environment.yaml --prune

Process the data.

bash scripts/run_process_data.sh  configs/openwebtext.json

Run the training and evaluation

Please see the run_train and run_eval scripts under the scripts directory. Example:

bash scripts/run_train.sh  configs/accelerate_1_gpu.yaml  configs/simple_data_test.json