JaydeepR/ldm-modernbert-base-sft

LDM-ModernBERT — Language Diffusion Model

A language diffusion model built on ModernBERT-base, pretrained on Project Gutenberg and fine-tuned on Open-Orca for instruction following.

Unlike autoregressive models that generate text left-to-right, this model generates text through iterative denoising — starting from a fully masked sequence and progressively unmasking tokens until a coherent output emerges.

[image]

Model Details

Training

Pretraining

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.

• —Dataset: Project Gutenberg (6,400,553 train chunks, 34,287 test chunks)
• —Initial train loss: 3.887 | Initial val loss: 3.922
• —Final train loss: 2.917 | Final val loss: 2.962

SFT (Supervised Fine-Tuning)

Fine-tuned on Open-Orca instruction-response pairs. Loss is computed only on the response tokens (not the instruction), using a query mask to identify answer boundaries.

• —Dataset: Open-Orca (~4.2M Q&A pairs)
• —Initial train loss: 1.559 | Initial val loss: 1.333
• —Final train loss: 0.837 | Final val loss: 0.967

Inference

The model supports two generation strategies:

• —`random` — masked tokens are randomly re-masked at each step
• —`low_confidence` — the lowest confidence tokens are re-masked, leading to more coherent outputs

Quickstart

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

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

Or use the provided inference scripts:

# Interactive inference
bash inference.sh

# Generate GIF
bash create_gif.sh

Limitations

• —Trained on a relatively small dataset (Project Gutenberg) with limited steps — quality is lower than production-scale models
• —SFT data was truncated to 1024 tokens; very long responses may be cut off
• —No RLHF or safety fine-tuning applied

Citation

Built following the approach from:

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