RFD4-Proteina 2026-10-19 release pair: flow-model training compute
Training compute of RFD4-Proteina 2026-10-19 release pair, native bundle
The preprint states the release flow model was pretrained with 256-token crops to 320,000 optimizer steps using four independently noised candidates per example, requiring about 16,000 GPU-hours, and then fine-tuned for about 2,000 GPU-hours with low-rank adaptation at rank 32, scaling factor 64 and no adapter dropout, over 60,000 additional updates with 384-token padding. Both flow stages use the same frozen autoencoder.
Descriptive fact about RFD4-Proteina 2026-10-19 release pair, taken from the pinned project sources named in the locator.
Evidence
Source checking verifies the cited claim or transcription. It does not establish independent reproduction.
Evidence table
Inspect claims, sources and review details
Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.
One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.
12 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Claimed field training_compute Context-only references | RFD4-Proteina Hugging Face model card at revision 621b785 Preprint version 3, Section S4.2 'Training stages' and Table S12 'Flow-model training stages' Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030 | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Archive member: artifacts/hf_README.md.gz |
| Claimed field training_compute Context-only references | Biomolecular design through latent reasoning with RFD4-Proteina, preprint version 3 Preprint version 3, Section S4.2 'Training stages' and Table S12 'Flow-model training stages' Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Version 3, uploaded 2026-11-13T07:04:35Z | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Source location Preprint version 3, Section S4.2 'Training stages' and Table S12 'Flow-model training stages' Context-only references | RFD4-Proteina Hugging Face model card at revision 621b785 Preprint version 3, Section S4.2 'Training stages' and Table S12 'Flow-model training stages' Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030 | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Archive member: artifacts/hf_README.md.gz |
| Source location Preprint version 3, Section S4.2 'Training stages' and Table S12 'Flow-model training stages' Context-only references | Biomolecular design through latent reasoning with RFD4-Proteina, preprint version 3 Preprint version 3, Section S4.2 'Training stages' and Table S12 'Flow-model training stages' Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Version 3, uploaded 2026-11-13T07:04:35Z | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Claimed value The preprint states the release flow model was pretrained with 256-token crops to 320,000 optimizer steps using four independently noised candidates per example, requiring about 16,000 GPU-hours, and then fine-tuned for about 2,000 GPU-hours with low-rank adaptation at rank 32, scaling factor 64 and no adapter dropout, over 60,000 additional updates with 384-token padding. Both flow stages use the same frozen autoencoder. Context-only references | RFD4-Proteina Hugging Face model card at revision 621b785 Preprint version 3, Section S4.2 'Training stages' and Table S12 'Flow-model training stages' Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030 | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Archive member: artifacts/hf_README.md.gz |
| Claimed value The preprint states the release flow model was pretrained with 256-token crops to 320,000 optimizer steps using four independently noised candidates per example, requiring about 16,000 GPU-hours, and then fine-tuned for about 2,000 GPU-hours with low-rank adaptation at rank 32, scaling factor 64 and no adapter dropout, over 60,000 additional updates with 384-token padding. Both flow stages use the same frozen autoencoder. Context-only references | Biomolecular design through latent reasoning with RFD4-Proteina, preprint version 3 Preprint version 3, Section S4.2 'Training stages' and Table S12 'Flow-model training stages' Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Version 3, uploaded 2026-11-13T07:04:35Z | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Description Descriptive fact about RFD4-Proteina 2026-10-19 release pair, taken from the pinned project sources named in the locator. Context-only references | RFD4-Proteina Hugging Face model card at revision 621b785 Preprint version 3, Section S4.2 'Training stages' and Table S12 'Flow-model training stages' Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030 | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Archive member: artifacts/hf_README.md.gz |
| Description Descriptive fact about RFD4-Proteina 2026-10-19 release pair, taken from the pinned project sources named in the locator. Context-only references | Biomolecular design through latent reasoning with RFD4-Proteina, preprint version 3 Preprint version 3, Section S4.2 'Training stages' and Table S12 'Flow-model training stages' Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Version 3, uploaded 2026-11-13T07:04:35Z | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Relationship: subject rfd4-proteina-20261115-config-release-2026-10-19 Context-only references | RFD4-Proteina Hugging Face model card at revision 621b785 Preprint version 3, Section S4.2 'Training stages' and Table S12 'Flow-model training stages' Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030 | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record Archive member: artifacts/hf_README.md.gz |
| Relationship: subject rfd4-proteina-20261115-config-release-2026-10-19 Context-only references | Biomolecular design through latent reasoning with RFD4-Proteina, preprint version 3 Preprint version 3, Section S4.2 'Training stages' and Table S12 'Flow-model training stages' Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: Version 3, uploaded 2026-11-13T07:04:35Z | not individually reviewed No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
Sources and history
Release 2026-11-21-c8bea25e5abb · Record review: source checked
2 source records and release history
- RFD4-Proteina Hugging Face model card at revision 621b785 · Original source · Hugging Face revision 621b785abf9d61dd24f4b81ae4dbd1330728f030
- Biomolecular design through latent reasoning with RFD4-Proteina, preprint version 3 · Original source · Version 3, uploaded 2026-11-13T07:04:35Z
Technical metadata and extraction receipts
Stable ID: rfd4-proteina-20261115-claim-flow-training
- areas
- proteins-complexes; protein-structure; molecular-interactions
- contexts
- research
- field
- training_compute
- review
- artifact sha256: 38479d8e4c6e46ed7e1ec4ece0dd02c358ae46d73d56790c44be813201fd9b98; method: independent-cell-check; source-hash-verification; pdf-text-parse; method note: Independent review pass. All thirteen pinned artifacts were re-retrieved on 2026-11-15 and every SHA-256 checked against its source record; the ten archived gzipped artifacts decompress to those bytes. Every sentence of this claim was read back against the locator it names in the re-retrieved artifact, and each statement attributed to a source was matched to that source's own wording. Preprint claims were checked against a pdftotext -layout text layer of the version 3 PDF retrieved at its versioned URL.; note: Every statement in this claim was found at the locator it names in the re-retrieved artifact. Checked against Section S4.2 and Table S12 of the re-retrieved version 3 PDF. Table S11 independently gives the release-scale (K = 4) total as about 18,000 GPU-hours, consistent with the 16,000 plus 2,000 this claim records.; retrieval url: https://huggingface.co/nvidia/RFD4-Proteina/raw/621b785abf9d61dd24f4b81ae4dbd1330728f030/README.md; reviewed at: 2026-11-15T00:00:00Z; reviewer: claude; reviewer note: Claude (Opus 5) review agent, independent of the agent that extracted this batch; no human review claimed. Transcription was checked against the pinned sources; nothing was reproduced and no value was recomputed.
- source locator
- Preprint version 3, Section S4.2 'Training stages' and Table S12 'Flow-model training stages'
- value
- The preprint states the release flow model was pretrained with 256-token crops to 320,000 optimizer steps using four independently noised candidates per example, requiring about 16,000 GPU-hours, and then fine-tuned for about 2,000 GPU-hours with low-rank adaptation at rank 32, scaling factor 64 and no adapter dropout, over 60,000 additional updates with 384-token padding. Both flow stages use the same frozen autoencoder.