YanZhanPKU/dLLM-PRM-Gap-Datasets

<h1 style="text-align: left; font-size: 1.6em; margin-bottom: 0.75em;"> <span style="color:#E69F00; font-weight:bold;">dLLM</span><span style="color:#0072B2; font-weight:bold;"> PRM</span><span style="color:#009E73; font-weight:bold;"> Gap</span> · Datasets </h1>

<div style="text-align: left; margin-bottom: 18px;"> <a href="https://github.com/dLLM-PRM-Gap/dLLM-PRM-Gap" target="blank">💻 Code</a> &nbsp; • &nbsp; <a href="https://huggingface.co/collections/YanZhanPKU/dllm-prm-gap" target="blank">🤗 Collection</a> &nbsp; • &nbsp; <a href="https://arxiv.org/abs/2609.35472" target="_blank">📄 Paper</a> </div>

This dataset repository contains the selected trajectory corpus and small diagnostic artifacts for dLLM PRM Gap, a controlled study of process reward model (PRM) guidance and outcome reward model (ORM) reranking in discrete diffusion reasoning.

Public release. This dataset accompanies the arXiv paper and the dLLM PRM Gap code release.

News

• —2026-09: Accepted at NeurIPS 2026; public corpus and diagnostic artifacts released.

Contents

The trajectory records are grouped by the release's source-generation conditions and retain the fields needed by the selected training and evaluation entry points. Source-machine paths are intentionally omitted from the published summary.

The main trajectory is a PyTorch artifact rather than a tabular dataset, so the Hub table viewer is not expected to render it. The benchmark task is derived from the public GSM8K task; base model weights and upstream benchmark packages are not redistributed here.

Quickstart

from pathlib import Path
import torch

path = Path("train_snapshots/trajectories_all.pt")
trajectories = torch.load(path, weights_only=False)
print(f"loaded {len(trajectories):,} trajectory records")
print(trajectories[0].keys())

Download the selected artifacts with the Hub CLI:

pip install -U "huggingface_hub[cli]"
huggingface-cli download YanZhanPKU/dLLM-PRM-Gap-Datasets \
  --repo-type dataset \
  --local-dir ./datasets

The release code loads the same file from datasets/train_snapshots/trajectories_all.pt. The JSON diagnostics can be inspected directly or regenerated with the corresponding scripts under `src/prm/`.

Relation to the code release

The companion repository provides:

• —PRM and ORM model definitions and checkpoint loading;
• —matched-compute denoising and candidate reranking entry points;
• —mask-ratio, readout, suffix, two-pass, top-M, MATH500, and LLaDA diagnostics;
• —focused protocol tests and the figures used by the paper release.

No full base-model, optimizer, private log, or machine-launch artifact is included in this dataset repository.

Provenance and license

• —The merged trajectory file is the selected GSM8K training/validation artifact used by the current paper baseline.
• —The selected model scope is summarized in the code release's `WEIGHT_PROVENANCE.md`.
• —The release code and packaging metadata are MIT licensed. GSM8K and the base models retain their upstream terms.

Citation

@misc{zhan2026deterministicprmguidanceunderperforms,
  title        = {Why Deterministic PRM Guidance Underperforms in Discrete Diffusion Reasoning},
  author       = {Yan Zhan and Shaobo Liu and Zhijun Gao},
  year         = {2026},
  eprint       = {2609.35472},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url          = {https://arxiv.org/abs/2609.35472},
}