| --- |
| 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. |
|
|
