README.md

---
license: apache-2.0
library_name: pytorch
tags:
- diffusion-models
- masked-diffusion
- token-ordering
- process-reward-model
---

# DPRM Release Artifacts

This bundle contains the retained DPRM states and raw evaluation records for
**DPRM: A Plug-in Token-Ordering Module for Diffusion Language Models**.
The executable integrations, exact commands, and uncertainty reducers are in
[DakeBU/DPRM-DLLM](https://github.com/DakeBU/DPRM-DLLM). The paper is available
at [arXiv:2604.24357](https://arxiv.org/abs/2604.24357), and the scrollable
visual galleries are on the [project page](https://dakebu.github.io/DPRM-DLLM/).

The bundle retains one deployable DPRM checkpoint or controller per host when
the host has trainable state. Baseline outputs are represented by raw matched
records and are regenerated with the public commands. `release_artifacts.json`
pins every reported file by byte size and SHA-256 digest.

| Host | Retained artifacts |
|---|---|
| puma | [dprm_2m](puma/dprm_confidence/ema_step_2000000.pt), [unmask3_question_traces](puma/traces/unmask3_all_1319.tar.zst) |
| dmpo | [math_confidence_records](dmpo/records/math_confidence_step5000.tar.gz), [math_dprm_records](dmpo/records/math_dprm_step5000.tar.gz), [countdown_dprm_checkpoint](dmpo/countdown/dprm_checkpoint_step5000.tar.gz), [countdown_confidence_records](dmpo/records/countdown_confidence_step5000.tar.gz), [countdown_dprm_records](dmpo/records/countdown_dprm_step5000.tar.gz), [matched_step5000_summary](dmpo/matched_step5000_summary.json), [record_source_map](dmpo/record_sources.json), [matched_step5000_rows](dmpo/matched_step5000_rows.csv) |
| prism | [confidence_records](prism/gsm8k/confidence_res.jsonl), [dprm_records](prism/gsm8k/dprm_res.jsonl) |
| dplm | [tchebycheff](dplm/dprm_tchebycheff/last.ckpt), [cogen200_records](dplm/records/cogen200_gate.tar.zst) |
| dcm | [recovery_controller](dcm/controllers/recovery.pt), [formal_cell_records](dcm/records/formal_cells_and_replications.tar.zst) |
| genmol | [balanced](genmol/preferences/balanced_5000.ckpt), [denovo_records](genmol/records/denovo_1000.tar.zst) |
| sdpo | [dprm_random](sdpo/dprm_random/model.pt), [formal_records](sdpo/records/formal_640_raw_records.tar.zst) |
| omni_diffusion | [selected_dprm_checkpoint_index](omni_diffusion/checkpoint-1000/model.safetensors.index.json), [selected_dprm_checkpoint_config](omni_diffusion/checkpoint-1000/config.json), [development_records](omni_diffusion/records/development128/records/two_encoder.json), [confirmation_records](omni_diffusion/records/confirmation512/records/two_encoder.json), [formal_result](omni_diffusion/online_action_value_release.json), [development_split](omni_diffusion/reproducibility/omni_partiprompts_development128.jsonl), [confirmation_split](omni_diffusion/reproducibility/omni_partiprompts_confirmation512.jsonl), [mechanism_manifest](omni_diffusion/mechanism_cases/manifest.json), [mechanism_figure](omni_diffusion/mechanism_cases/paper_figure.png), [checkpoint_model_00001_of_00004](omni_diffusion/checkpoint-1000/model-00001-of-00004.safetensors), [checkpoint_model_00002_of_00004](omni_diffusion/checkpoint-1000/model-00002-of-00004.safetensors), [checkpoint_model_00003_of_00004](omni_diffusion/checkpoint-1000/model-00003-of-00004.safetensors), [checkpoint_model_00004_of_00004](omni_diffusion/checkpoint-1000/model-00004-of-00004.safetensors), [supplement_mechanism_manifest](omni_diffusion/supplementary_mechanism_cases/case_manifest.json), [supplement_replay_manifest](omni_diffusion/supplementary_mechanism_cases/replay_manifest.json), [supplement_replay_archive](omni_diffusion/supplementary_mechanism_cases.tar.gz), [supplement_omni_mechanism_cases_1](omni_diffusion/supplementary_mechanism_cases/figures/omni_mechanism_cases_1.png), [supplement_omni_mechanism_cases_2](omni_diffusion/supplementary_mechanism_cases/figures/omni_mechanism_cases_2.png), [supplement_omni_mechanism_cases_3](omni_diffusion/supplementary_mechanism_cases/figures/omni_mechanism_cases_3.png), [supplement_omni_mechanism_cases_4](omni_diffusion/supplementary_mechanism_cases/figures/omni_mechanism_cases_4.png), [qualitative_gallery_manifest](omni_diffusion/qualitative_gallery/omni_qualitative_gallery.json), [qualitative_gallery_records](omni_diffusion/qualitative_gallery/omni_qualitative_gallery_records.json), [qualitative_gallery_1](omni_diffusion/qualitative_gallery/omni_qualitative_gallery_1.png), [qualitative_gallery_2](omni_diffusion/qualitative_gallery/omni_qualitative_gallery_2.png), [qualitative_gallery_3](omni_diffusion/qualitative_gallery/omni_qualitative_gallery_3.png), [qualitative_gallery_4](omni_diffusion/qualitative_gallery/omni_qualitative_gallery_4.png) |
| llada_v | [rwqa_controller](llada_v/controllers/p1_b8_pos4.json), [rwqa_records](llada_v/records/realworldqa_split_records.tar.zst), [ai2d_preregistered_diagnostic](llada_v/diagnostics/ai2d_preregistered_confirmation.tar.gz), [qualitative_gallery_manifest](llada_v/qualitative_gallery/llada_v_qualitative_gallery.json), [qualitative_gallery_records](llada_v/qualitative_gallery/lladav_realworldqa_gallery_records.json), [qualitative_gallery_1](llada_v/qualitative_gallery/lladav_realworldqa_gallery_1.png), [qualitative_gallery_2](llada_v/qualitative_gallery/lladav_realworldqa_gallery_2.png) |

## Verification

```bash
git clone https://github.com/DakeBU/DPRM-DLLM.git
cd DPRM-DLLM
python scripts/verify_artifact_manifest.py \
  --manifest "$DPRM_ARTIFACT_ROOT/release_artifacts.json" \
  --artifact-root "$DPRM_ARTIFACT_ROOT" --require-complete
python scripts/audit_artifact_semantics.py \
  --artifact-root "$DPRM_ARTIFACT_ROOT"
```

See each integration README in the code repository for upstream model and
dataset requirements. These artifacts do not replace the upstream licenses.