README.md

---
datasets:
- Skylion007/openwebtext
papers:
- arxiv: 2604.11748
language:
- en
library_name: transformers
license: apache-2.0
metrics:
- perplexity
pipeline_tag: text-generation
---

# LangFlow

LangFlow is a continuous diffusion language model that operates in embedding space. Unlike discrete diffusion models (MDLM, SEDD, DUO), LangFlow performs diffusion directly on continuous token embeddings, enabling smoother denoising dynamics.

For more details, please see our paper: [LangFlow: Continuous Diffusion Rivals Discrete in Language Modeling](https://arxiv.org/abs/2604.11748).


## Using LangFlow

To use the pre-trained model for text generation, use the following snippet:

```python
from transformers import AutoModelForMaskedLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained('gpt2')
model = AutoModelForMaskedLM.from_pretrained('chumengl/langflow-owt', trust_remote_code=True)

# Generate samples
samples = model.generate_samples(num_samples=5, num_steps=128)
texts = tokenizer.batch_decode(samples, skip_special_tokens=True)
for text in texts:
    print(text)
```

## Model Details

- **Architecture**: DiT (Diffusion Transformer) backbone with adaptive layer normalization
- **Context Length**: 1024 tokens
- **Parameters**: ~130M non-embedding parameters (similar to GPT-2 medium)
- **Training**: 1M steps on OpenWebText corpus
- **Tokenizer**: GPT-2 tokenizer (50,257 vocab size)

## Citation

```
@article{chen2026langflow,
  title={LangFlow: Continuous Diffusion Rivals Discrete in Language Modeling},
  author={Chen, Yuxin and Liang, Chumeng and Sui, Hangke and Guo, Ruihan and Cheng, Chaoran and You, Jiaxuan and Liu, Ge},
  journal={arXiv preprint arXiv:2604.11748},
  year={2026}
}
```

## Model Card Contact

Chumeng Liang (chumengl@illinois.edu)