ESM-DCLM-1B

OpenESM 1B model (variant d26) trained on DCLM. This repository contains a Hugging Face Transformers-compatible export for the pretraining checkpoint. It uses OpenESM's custom architecture and remote model code; it is not the built-in transformers.EsmModel architecture.

How to Use

Install the runtime dependencies:

pip install torch transformers

Load the tokenizer and model with the custom OpenESM code enabled:

import torch
from transformers import AutoModelForMaskedLM, AutoTokenizer

repo_id = "guan-wang/ESM-DCLM-1B"
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForMaskedLM.from_pretrained(
    repo_id,
    trust_remote_code=True,
)

device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
inputs = tokenizer("Hello world", return_tensors="pt").to(device)
with torch.inference_mode():
    outputs = model(**inputs)

print(outputs.logits.shape)

The first load may prompt you to review and allow the repository's custom model code. Loading remote Python code requires trust_remote_code=True.

Model Details

Detail	Value
Model family	OpenESM energy-based language model
Architecture	Custom OpenESM implementation (d26)
Parameter scale	1B
Training dataset	DCLM
Training stage	pretraining
Transformer blocks	26
Embedding dimension	1664
Attention heads	13
Context length	2048
Vocabulary size	32768

The parameter scale is the label used for this model in the OpenESM model configuration. The architecture and sequence settings above are read from the exported checkpoint configuration.

Files

• config.json: model configuration.
• model*.safetensors: model weights in the standard Transformers format.
• modeling_esm.py: remote model and tokenizer implementation.
• configuration_esm.py: remote configuration implementation.
• tokenizer.pkl: serialized ESM tokenizer.
• tokenizer_config.json: tokenizer auto-loading configuration.
• token_bytes.pt: token byte table used by OpenESM metrics.

The original Lightning .ckpt is converted before upload and is not needed to load this repository. Training token counts are intentionally omitted from the repository name and file names.

Checkpoint Metadata

ebm-dclm-d26-10b

EBM pretrained on DCLM; d26, 10B.

This repository contains the following PyTorch checkpoint:

• Archive filename: final-s=step=9499-d26-ctx2048-lr0.0012-bs1x32-muon_adamw-valid_loss=valid_loss=2.4950.ckpt
• Original filename: s=step=9499-d26-ctx2048-lr0.0012-bs1x32-muon_adamw-valid_loss=valid_loss=2.4950.ckpt
• Original relative path: pretrain/ebm/dclm/d26/10B/s=step=9499-d26-ctx2048-lr0.0012-bs1x32-muon_adamw-valid_loss=valid_loss=2.4950.ckpt

The checkpoint is uploaded as-is from the local training output. See the filename and training project for the exact architecture and loading code.

License

Please add the applicable model/data license before publishing this repo.

The code is maintained in the OpenESM repository.