zpointsun / DiBO-TFBind8

DiBO-TFBind8

Final task-specific DiBO model for TFBind8-Exact-v0 , released with Training Diffusion Language Models for Black-Box Optimization (ICML 2026 Spotlight). See also the Hugging Face paper page and the DiBO code repository .

This model completed domain adaptation (DA), supervised fine-tuning (SFT), and reinforcement learning (RL).

Available model formats

This repository provides the same final task-specific DiBO model in two formats.

1. Original PyTorch checkpoint. dibo_tfbind8_final.pt is the canonical paper-faithful checkpoint produced by the DiBO training pipeline. It stores the state dictionary under the model key and is loaded through the DiBO codebase on top of the pinned LLaDA base revision.
2. Transformers/safetensors export. The root-level config, tokenizer, custom modeling code, and sharded safetensors files are a validated convenience export derived deterministically from the original checkpoint. They load directly with AutoModel.from_pretrained(...) .

The safetensors model was not trained separately. The LLaDA base weights are not duplicated in this repository.

Artifact details

• Base model: GSAI-ML/LLaDA-8B-Instruct , revision 08b83a6feb34df1a6011b80c3c00c7563e963b07
• Original checkpoint: dibo_tfbind8_final.pt (16,030,379,937 bytes; SHA-256 0d8e17f7aa60baf4173d71a4ec8475d182a992812b59d29a728eec719d1b67bf )
• Original format: PyTorch checkpoint dictionary, state dict at model
• Transformers model class: DiBOLLaDAModelLM loaded with AutoModel
• Tokenizer: PreTrainedTokenizerFast , vocabulary size 126,353
• Input/output vocabulary rows: 126,353 / 126,464
• Required delimiter tokens: |design-start| , |design-end| , |label-start| , |label-end|

The standard export contains the following safetensors shards:

• model-00001-of-00004.safetensors (4,994,672,360 bytes; SHA-256 b1d0d9d09be47f1c47dff91926fd7f3eda272dc2ecbdd1b1a68f4bf42984473c )
• model-00002-of-00004.safetensors (4,999,819,552 bytes; SHA-256 dbe8101a92f660a2c2bd3d5d3dae6cc58f8258ccafc5d487d627f4aa7f48f7ad )
• model-00003-of-00004.safetensors (4,999,802,728 bytes; SHA-256 62c12ec9f508f42811eaa1880dfcb7c4fbca1a48d8379e796d2e8f34f978abfe )
• model-00004-of-00004.safetensors (1,035,993,232 bytes; SHA-256 0a6587d94453234ffa5d0e18cddfde93219c40754a2d5359e0ce1e095f714286 )

Its total weight size is 16,030,287,872 bytes. The machine-readable artifact manifest and validation report record source hashes and measured validation results. Validation passed for: candidate parity, config equality, forward parity, remote from pretrained, remote pt download, strict load, task smoke, tokenizer equality, weight equality.

A. Load the standard Transformers export

from transformers import AutoModel, AutoTokenizer

repo_id = "zpointsun/DiBO-TFBind8"
tokenizer = AutoTokenizer.from_pretrained(
    repo_id,
    revision="v1.1.1",
    trust_remote_code=True,
)
model = AutoModel.from_pretrained(
    repo_id,
    revision="v1.1.1",
    trust_remote_code=True,
    use_safetensors=True,
    torch_dtype="auto",
)
model.eval()

The packaged tokenizer already includes the four DiBO delimiter tokens. Do not add them or resize embeddings again after loading this export.

B. Download and load the original checkpoint

The original artifact uses the released DiBO loader, which initializes the pinned LLaDA base, adds the four delimiter tokens, resizes the input embedding, and strictly loads checkpoint["model"] .

hf download zpointsun/DiBO-TFBind8 dibo_tfbind8_final.pt \
  --revision v1.1.1 --local-dir checkpoints/dibo-tfbind8

import torch
from huggingface_hub import hf_hub_download
from src.model.dllm import DEFAULT_MODEL_ID, LLADA_MODEL_REVISION, load_model_and_tokenizer

assert DEFAULT_MODEL_ID == "GSAI-ML/LLaDA-8B-Instruct"
assert LLADA_MODEL_REVISION == "08b83a6feb34df1a6011b80c3c00c7563e963b07"
checkpoint_path = hf_hub_download(
    "zpointsun/DiBO-TFBind8",
    filename="dibo_tfbind8_final.pt",
    revision="v1.1.1",
)
model, tokenizer = load_model_and_tokenizer(DEFAULT_MODEL_ID, device="cuda")
checkpoint = torch.load(checkpoint_path, map_location="cuda")
model.load_state_dict(checkpoint["model"], strict=True)
model.eval()

C. Evaluate either format

From a checkout of the released DiBO code and its oracle environment:

# Standard Transformers export
python eval.py --tasks TFBind8-Exact-v0 \
  --model_name_or_path zpointsun/DiBO-TFBind8 --model_revision v1.1.1 \
  --seeds <SEEDS> --max_attempts 1000

# Canonical local .pt checkpoint
python eval.py --tasks TFBind8-Exact-v0 \
  --checkpoint_path checkpoints/dibo-tfbind8/dibo_tfbind8_final.pt \
  --seeds <SEEDS> --max_attempts 1000

Both choices share the same downstream DiBO evaluation path. Direct oracle evaluation requires the Design-Bench data cache and task dependencies described in the DiBO repository . For the exact Design-Bench snapshot used in the DiBO experiments, see DiBO-DesignBench-Snapshot .

Limitations

Practical inference requires a CUDA-capable PyTorch environment. These task-specific models are designed for DiBO's masked-response generation and evaluation workflow; this release does not claim generic text-generation pipeline support. Loading a released final model is for evaluation or use and does not reproduce the DA/SFT/RL training process.

Other DiBO task models

• DiBO-TFBind10
• DiBO-AntMorphology
• DiBO-DKittyMorphology

Citation

If you find DiBO helpful, please cite:

@article{sun2026training,
  title={Training diffusion language models for black-box optimization},
  author={Sun, Zipeng and Chen, Can and Yuan, Ye and Wu, Haolun and Gu, Jiayao and Pal, Christopher and Liu, Xue},
  journal={arXiv preprint arXiv:2603.17919},
  year={2026}
}