JaydeepR/ldm-modernbert-base-pretrain

LDM-ModernBERT — Pretrained Language Diffusion Model

A language diffusion model built on ModernBERT-base, pretrained on Project Gutenberg using a masked diffusion objective.

This is the base pretrained checkpoint before SFT instruction tuning. For instruction following, see JaydeepR/ldm-modernbert-base-sft.

[image]

Model Details

Training

The model is pretrained using a flow-matching diffusion objective: at each step, a random fraction t of tokens is masked, and the model learns to predict the original tokens. The loss is scaled by 1/t to account for the difficulty of predicting heavily masked sequences.

Inference

from transformers import AutoModelForMaskedLM
from safetensors.torch import load_file
import torch

model = AutoModelForMaskedLM.from_pretrained("answerdotai/ModernBERT-base")
state_dict = load_file("model.safetensors")
model.load_state_dict(state_dict, strict=False)
model.eval()

# Unconditional generation — start from all masked tokens
seq_len = 128
input_tokens = torch.full((1, seq_len), tokenizer.mask_token_id, dtype=torch.long)

Or use the provided scripts from the GitHub repo:

# Generate GIF (unconditional)
bash create_gif.sh

Limitations

• —Trained on a relatively small dataset (Project Gutenberg) with limited steps
• —No instruction tuning — use the SFT checkpoint for Q&A tasks
• —Output has a literary/formal style reflecting Gutenberg training data

Citation

Built following the approach from:

• —Masked Diffusion Language Models
• —PyTorch-Adventures — Language Diffusion Model by @priyammaz