File size: 1,255 Bytes
e12a518
 
 
 
 
 
 
 
 
7e6b4e2
 
e12a518
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7e6b4e2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
---
license: mit
tags:
- diffusion
- llm
- conversational
- difference-labs
datasets:
- smangrul/ultrachat-10k-chatml
base_model:
- darwinkernelpanic/DiffReaper-5L
---

# DiffReaper-6

**DiffReaper-6** is a Large-scale Diffusion-based Large Language Model (Diffusion-LLM) developed by **DifferenceLabs**. 

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.

## Model Details
- **Architecture**: Diffusion-Transformer (DiT) with Adaptive Layer Norm (adaLN-Single) modulation.
- **Backbone**: 24 Layers, 24 Attention Heads, 1536 Hidden Dimension.
- **Tokenizer**: BERT-base-uncased.
- **Training Objective**: MSE on Denoising Latents (Predicting original embeddings from noisy input).
- **Conditioning**: Prompt-concatenated latents with time-step embedding.

## Training
The model is being trained on an RTX 5090 using the `ultrachat-10k` dataset, focusing on conversational flow and instruction following.

## Goal
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.