Upload README.md (0.0 MB)

README.md

@@ -0,0 +1,33 @@
+---
+license: apache-2.0
+tags:
+- flow-matching
+- continuous-latents
+- math-reasoning
+- qwen2.5
+- block-diffusion
+size_categories:
+- 10K<n<100K
+task_categories:
+- text-generation
+---
+
+# BlockDiffuse Precomputed Latents & Reasoning Datasets
+
+Precomputed tokenized reasoning traces and continuous latent trajectory representations extracted from **Qwen/Qwen2.5-0.5B-Instruct** for training **BlockDiffuse** Diffusion Transformers via Rectified Flow Matching.
+
+## Contents
+- `reasoning_tokenized_qwen.pt`: Pre-tokenized reasoning sequences formatted for block-wise training (12.5 MB).
+- `precomputed_reasoning_latents_qwen.pt`: Continuous latent representations extracted from intermediate layers of Qwen2.5 (69.1 MB).
+- `precomputed_real_qwen.pt`: Extracted mid-layer latent vectors and velocity targets (1,045 MB).
+- `precomputed_real_qwen_full.pt`: Full-scale training set with multi-turn reasoning traces (2,360 MB).
+
+## Citation
+```bibtex
+@article{blockdiffuse2026,
+  title={BlockDiffuse: Fully Parallel Latent Space Reasoning Generation with Diffusion Transformers},
+  author={Hooshaai Research},
+  year={2026},
+  url={https://github.com/Hooshaai/BlockDiffuse}
+}
+```