| --- |
| license: apache-2.0 |
| tags: |
| - flow-matching |
| - continuous-latents |
| - math-reasoning |
| - qwen2.5 |
| - block-diffusion |
| - non-autoregressive |
| size_categories: |
| - 10K<n<100K |
| task_categories: |
| - text-generation |
| language: |
| - en |
| --- |
| |
| # 📦 BlockDiffuse Precomputed Latents & Reasoning Datasets |
|
|
| [](https://opensource.org/licenses/Apache-2.0) |
| [](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) |
| [](https://huggingface.co/tahamajs/BlockDiffuse) |
| [](https://huggingface.co/spaces/tahamajs/BlockDiffuse-Blog) |
|
|
| This repository contains the complete pre-extracted dataset files used to train **BlockDiffuse** Diffusion Transformers to generate **100-token blocks in continuous latent space** using **Rectified Flow Matching**. |
|
|
| --- |
|
|
| ## 🔬 Dataset Overview & Extraction Pipeline |
|
|
| Standard language models operate over discrete token vocabularies ($V \approx 151{,}936$). To bypass sequential autoregressive decoding, BlockDiffuse maps prompts and target answer sequences into continuous latent vectors: |
|
|
| ``` |
| Discrete Prompt Tokens (L_p) ──► Qwen2.5-0.5B (Layer 12) ──► Prompt Latents c [L_p x 896] |
| Discrete Target Tokens (100) ──► Qwen2.5-0.5B (Layer 12) ──► Target Latents z_1 [100 x 896] |
| ``` |
|
|
| These precomputed continuous tensors allow training the Diffusion Transformer directly on latent trajectory matching without re-computing LLM forward passes on every iteration, accelerating training throughput by **> 12x**. |
|
|
| --- |
|
|
| ## 📁 File Manifest & Specifications |
|
|
| | File Name | File Size | Description | Shape / Keys | |
| | :--- | :--- | :--- | :--- | |
| | `reasoning_tokenized_qwen.pt` | **12.5 MB** | Pre-tokenized GSM8K & Math reasoning conversations formatted with the Qwen2.5 ChatML template (`<\|im_start\|>...<\|im_end\|>`). | `input_ids`, `attention_mask`, `labels` | |
| | `precomputed_reasoning_latents_qwen.pt` | **69.1 MB** | Validation subset of continuous target latents ($z_1$) and prompt conditionings ($c$) extracted from Layer 12 of Qwen2.5. | `{"prompt_latents": [N, L_p, 896], "target_latents": [N, 100, 896]}` | |
| | `precomputed_real_qwen.pt` | **1,045.9 MB** | Intermediate-scale latent training dataset containing 1,000 multi-turn mathematical reasoning trajectories. | `{"prompt_latents", "target_latents", "target_tokens"}` | |
| | `precomputed_real_qwen_full.pt` | **2,360.3 MB** | Complete production training dataset covering extensive multi-step reasoning problems. | `{"prompt_latents", "target_latents", "target_tokens"}` | |
|
|
| --- |
|
|
| ## 🚀 How to Load and Use |
|
|
| ### 1. Direct Python Loading via `torch.load` |
| ```python |
| import torch |
| |
| # Load tokenized sequences |
| tokenized_data = torch.load("reasoning_tokenized_qwen.pt", map_location="cpu") |
| print("Tokenized sample count:", len(tokenized_data["input_ids"])) |
| |
| # Load precomputed continuous latents |
| latents_data = torch.load("precomputed_reasoning_latents_qwen.pt", map_location="cpu") |
| print("Prompt latents shape:", latents_data["prompt_latents"][0].shape) # [L_p, 896] |
| print("Target latents shape:", latents_data["target_latents"][0].shape) # [100, 896] |
| ``` |
|
|
| ### 2. Training BlockDiffuse DiT with this Dataset |
| ```bash |
| # Clone official codebase |
| git clone https://github.com/Hooshaai/BlockDiffuse.git |
| cd BlockDiffuse |
| |
| # Train with the precomputed full dataset |
| python train.py \ |
| --config_train configs/gpu_full_capacity_improved.yaml \ |
| --config_dit configs/gpu_full_capacity_improved.yaml \ |
| --data_path ./data/precomputed_real_qwen_full.pt \ |
| --max_steps 20000 \ |
| --output_dir ./checkpoints_improved |
| ``` |
|
|
| --- |
|
|
| ## 📐 Latent Space Normalization & Properties |
| - **Dimensionality**: $d_{\text{model}} = 896$ per token position. |
| - **Layer Origin**: Extracted after RMSNorm from Transformer Block 12 of `Qwen2.5-0.5B-Instruct`. |
| - **Target Block Length**: Exactly 100 contiguous tokens. Shorter sequences are padded to 100 with EOS token latents; longer reasoning traces are chunked with rolling context propagation. |
| |
| --- |
| |
| ## 📜 Citation |
| ```bibtex |
| @article{blockdiffuse2026, |
| title={BlockDiffuse: Fully Parallel Latent Space Reasoning Generation with Diffusion Transformers}, |
| author={Hooshaai Research}, |
| journal={GitHub / HuggingFace Technical Report}, |
| year={2026}, |
| url={https://github.com/Hooshaai/BlockDiffuse} |
| } |
| ``` |
| |
