WilhelmH's picture
Bridge-7b-Diffusion-Full: Epoch 1 checkpoint (val_loss=0.1915)
7a36681 verified
metadata
license: apache-2.0
base_model: Dream-org/Dream-v0-Instruct-7B
tags:
  - diffusion
  - reasoning
  - reversethought
  - dream
datasets:
  - ianncity/KIMI-K2.5-1000000x
pipeline_tag: text-generation

Bridge-7b-Diffusion

A fine-tuned DREAM 7B masked diffusion language model trained with the ReverseThought objective.

What is ReverseThought?

Given a question and its answer, the model learns to produce the step-by-step reasoning chain that bridges the question to the answer. This trains the model to generate coherent chain-of-thought reasoning via DREAM's masked diffusion process.

  • Input: Question + Answer
  • Output: Detailed reasoning trace connecting them

Training Details

  • Base model: Dream-org/Dream-v0-Instruct-7B
  • Training data: 75,000 examples from KIMI-K2.5-1000000x (General-Distillation subset)
  • Objective: DREAM masked diffusion with CART time reweighting
  • Hardware: 8x NVIDIA H100 80GB
  • Epochs: 3
  • Batch size: 128
  • Learning rate: 2e-6 (cosine schedule)
  • Max sequence length: 2048 tokens
  • Precision: bf16 mixed precision (FSDP)

Usage

from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("WilhelmH/Bridge-7b-Diffusion", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("WilhelmH/Bridge-7b-Diffusion", trust_remote_code=True)

Architecture

This is a masked diffusion language model (not autoregressive). It uses bidirectional attention and generates text by iteratively denoising masked tokens. See the DREAM paper for details.