dataset_info:
features:
- name: text_trg
dtype: string
splits:
- name: train
num_bytes: 16486231249
num_examples: 36074196
- name: test
num_bytes: 22761845
num_examples: 50000
download_size: 11310715074
dataset_size: 16508993094
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
task_categories:
- text-generation
language:
- en
This dataset is associated with the paper Compressed and Smooth Latent Space for Text Diffusion Modeling, which introduces COSMOS, a novel approach to text generation that operates entirely in a compressed, smooth latent space tailored specifically for diffusion models. This space is learned using an autoencoder trained simultaneously for token-level reconstruction and alignment with frozen activations from a pretrained language encoder, providing robust semantic grounding.
The dataset, featuring text_trg (target text), is used to train and evaluate models within the COSMOS framework across diverse generative tasks, including story generation, question generation, summarization, and detoxification. The paper demonstrates that text representations can be compressed by $8\times$ while maintaining generation quality comparable to token-level diffusion models, and can surpass both diffusion-based and autoregressive baselines.
- Paper: Compressed and Smooth Latent Space for Text Diffusion Modeling
- Code: https://github.com/MeshchaninovViacheslav/cosmos
Paper Abstract
Autoregressive language models dominate modern text generation, yet their sequential nature introduces fundamental limitations: decoding is slow, and maintaining global coherence remains challenging. Diffusion models offer a promising alternative by enabling parallel generation and flexible control; however, their application to text generation is hindered by the high dimensionality of token-level representations. We introduce Cosmos, a novel approach to text generation that operates entirely in a compressed, smooth latent space tailored specifically for diffusion. This space is learned using an autoencoder trained simultaneously for token-level reconstruction and alignment with frozen activations from a pretrained language encoder, providing robust semantic grounding and enabling effective perturbation-based augmentations. Empirically, we demonstrate that text representations can be compressed by $8\times$ while maintaining generation quality comparable to token-level diffusion models. Furthermore, increasing the latent sequence length allows Cosmos to surpass both diffusion-based and autoregressive baselines. We evaluate Cosmos on four diverse generative tasks including story generation, question generation, summarization, and detoxification and compare it with various generative paradigms. Cosmos achieves comparable or superior generation quality while offering more than $2\times$ faster inference. Code is released at \href{ this https URL }{GitHub}
Sample Usage
After training the diffusion model, you can generate new text samples using the following command, as described in the project's GitHub repository:
CUDA_LAUNCH_BLOCKING=1 \
HYDRA_FULL_ERROR=1 \
uv run \
torchrun --nproc_per_node=4 --master_port=12345 \
generate.py \
dataset=rocstories \
diffusion.dynamic.N=200 \
diffusion.dynamic.d=5 \
diffusion.training.batch_size=512 \
encoder.latent.num_latents=16 \
encoder.embedding.max_position_embeddings=128 \
decoder.latent.num_latents=16 \
decoder.embedding.max_position_embeddings=128 \
autoencoder.model.load_checkpoint='"autoencoder-num_latents=16-wikipedia-final-128/100000.pth"' \
diffusion.model.load_checkpoint='"diffusion-rocstories-16-d=5-final/180000.pth"' \
diffusion.generation.num_gen_texts=2000 \
training=""
Citation
If you use this work or the associated dataset, please cite the paper:
@misc{meshchaninov2025compressed,
title={Compressed and Smooth Latent Space for Text Diffusion Modeling},
author={Viacheslav Meshchaninov and Egor Chimbulatov and Alexander Shabalin and Aleksandr Abramov and Dmitry Vetrov},
year={2025},
eprint={2506.21170},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2506.21170},
}