Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
library_name: diffusers
|
| 4 |
+
tags:
|
| 5 |
+
- diffusion
|
| 6 |
+
- llm
|
| 7 |
+
- conversational
|
| 8 |
+
- difference-labs
|
| 9 |
+
datasets:
|
| 10 |
+
- smangrul/ultrachat-10k-chatml
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# DiffReaper-6
|
| 14 |
+
|
| 15 |
+
**DiffReaper-6** is a Large-scale Diffusion-based Large Language Model (Diffusion-LLM) developed by **DifferenceLabs**.
|
| 16 |
+
|
| 17 |
+
It represents a significant architectural leap over the previous 5L version, transitioning to a more robust denoiser and a deeper transformer-based backbone to achieve actual conversational coherence.
|
| 18 |
+
|
| 19 |
+
## Model Details
|
| 20 |
+
- **Architecture**: Diffusion-Transformer (DiT) with Adaptive Layer Norm (adaLN-Single) modulation.
|
| 21 |
+
- **Backbone**: 24 Layers, 24 Attention Heads, 1536 Hidden Dimension.
|
| 22 |
+
- **Tokenizer**: BERT-base-uncased.
|
| 23 |
+
- **Training Objective**: MSE on Denoising Latents (Predicting original embeddings from noisy input).
|
| 24 |
+
- **Conditioning**: Prompt-concatenated latents with time-step embedding.
|
| 25 |
+
|
| 26 |
+
## Training
|
| 27 |
+
The model is being trained on an RTX 5090 using the `ultrachat-10k` dataset, focusing on conversational flow and instruction following.
|
| 28 |
+
|
| 29 |
+
## Goal
|
| 30 |
+
To prove that diffusion models can reach (and eventually exceed) the coherence of auto-regressive models while maintaining the creative "soul" and parallel generation benefits of diffusion.
|