Riddle Diffusion Phase3

Riddle Diffusion Phase3 is a text generation model designed in the Diffusion-LM style, following the approach described by Li et al. (2022). It is trained specifically to address the task of solving synthetic riddles, using a dataset of 232 deduplicated synthetic riddles known as "Riddles — A Synthetic Riddle Dataset for NLP" (licensed CC0). The model operates by taking a riddle as context and then iteratively denoising Gaussian noise into answer word embeddings through a learned reverse process, leveraging continuous embedding diffusion combined with a Transformer encoder/decoder architecture. The model contains 8,024,576 parameters, with a configuration that includes 200 timesteps, a model dimension of 256, four layers, a feedforward dimension of 1024, and four attention heads. Its performance on held-out data is reported as achieving an exact match rate of 47.0% with K=1 and K=10. The intended use of Riddle Diffusion Phase3 is for research and diagnostics, and it is explicitly noted as not being a production-ready riddle solver. Delivery is through the Hugging Face platform, and the model is provided under the MIT license. The evidence does not specify any integrations, deployment formats beyond the Hugging Face model card, or additional user interface details. The tool is positioned as a research artifact within the class of diffusion-based language models for text generation, with a focus on riddle-solving tasks in natural language processing.