EditJumps released model card on Hugging Face
Repository revision 44a0ca3e224e16adac0e15e24400d22a17d4db5c, last modified 2026-09-23 · Project source
What this source contributed
Records in this release that cite this source. A record can cite several sources.
Claims
Statements this source supports, each with where in the source it was checked.
- Best validation loss differs between the paper and the model card: Validation loss The model card gives a best validation loss of 130.48 at step 18,500 and marks the step with its own 'TODO: confirm step' comment, while Appendix D of the paper gives 135.2 at step 19,999 and no value below it. The two figures are not reconciled in either source. (Model card 'Model details', Best val loss row; paper Appendix D 'Training Protocol and Compute Resources')
- Code and weights licence: Licence All code, model weights and evaluation pipelines for EditJumps are released under an MIT licence. The repository and the model repository serve byte-identical MIT licence files, and the model card states MIT for both the code and the weights. (Appendix D 'Operational Integrity and Artifact Verification'; repository LICENSE; model card 'License')
- Released checkpoint parameter count: Parameters 35,213,590 (Model card 'Model details', Parameters row)
- Released checkpoint training budget: Training budget 20,000 steps at batch size 16, which the card gives as 320,000 pairs and about 19% of one epoch, and which it states is not converged and not claimed to be. Adam, learning rate 1e-4 constant, no warm-up, decay or weight decay, gradient clipping 1.0, seed 0. (Model card 'Model details', Budget and Optimizer rows)
- Released checkpoint training corpus: Training data 1,660,105 unordered homolog pairs from the Observed Antibody Space paired corpus of 2,586,927 VH.VL sequences. Clusters were built with MMseqs2 easy-cluster at minimum sequence identity 0.5 and coverage 0.8 and capped at 20 pairs per family. (Model card 'Model details', Training data row; paper Appendix D 'Corpus Curation and Homolog Clustering')
- Released checkpoint trunk identity: Trunk Stock facebook/esm2_t12_35M_UR50D: 12 hidden layers, hidden size 480, intermediate size 1920, 20 attention heads, vocabulary 33, rotary position embeddings, fp32. The card states the trunk is deliberately stock ESM-2 rather than the project's own OAS-adapted trunk, because the EvoFlows specification calls for no domain adaptation. (Model card 'Model details', Trunk and Precision rows, and the paragraph below the table; model/encoder/config.json)
- The scored arm and the released weights state different head parameterisations: Head Section 3 of the paper lists 'linear rather than MLP rate heads' as one of four operational adaptations of the scored arm, while the model card's Heads row gives three MLP rate heads and its quickstart loads the released weights with rate_head mlp and q_head esm_lm_head. Appendix E reports the two parameterisations give different edit ratios, 0.19 for the published Appendix A heads against 0.22 to 0.28 for lightweight linear heads, so they are not interchangeable. (Section 3 'Supervision discrepancies and deviations'; Figure 3 caption; Appendix E 'Non-Explanations for the Edit Budget Discrepancy'; model card 'Model details', Heads row and the quickstart command)
- Training runtime as the model card states it: Runtime The model card's Training cost row gives about 3 h on one A100 at about 0.54 s/step and about 5.8 h on an L4. (Model card 'Model details', Training cost row)
- What the licence does not cover: Licence scope The MIT grant covers this project's code and released weights. It does not cover the pretrained ESM-2 checkpoint the editor is fine-tuned from or the Observed Antibody Space corpus it is trained on, which carry their own separate terms and have their own source records. (Model card 'Model details', Trunk and Training data rows, and 'License'; repository LICENSE)
Models, benchmarks and other records
Methods (1)
Configurations (1)
Datasets (2)
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.
28 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Access and reuse The card declares MIT for the code and the weights. The stock ESM-2 trunk it builds on and the Observed Antibody Space corpus it is trained from carry their own separate terms. Source metadata | EditJumps released model card on Hugging Face No field-specific location recorded Version: Repository revision 44a0ca3e224e16adac0e15e24400d22a17d4db5c, last modified 2026-09-23 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Artifact SHA-256 552199e1e3be6d07de661baf23876fcbc354f8cfd33b391d2913291a3a5a8c32 Source metadata | EditJumps released model card on Hugging Face No field-specific location recorded Version: Repository revision 44a0ca3e224e16adac0e15e24400d22a17d4db5c, last modified 2026-09-23 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Artifact url https://huggingface.co/VisiumSA/EditJumps/raw/44a0ca3e224e16adac0e15e24400d22a17d4db5c/README.md Source metadata | EditJumps released model card on Hugging Face No field-specific location recorded Version: Repository revision 44a0ca3e224e16adac0e15e24400d22a17d4db5c, last modified 2026-09-23 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Licence MIT Source metadata | EditJumps released model card on Hugging Face No field-specific location recorded Version: Repository revision 44a0ca3e224e16adac0e15e24400d22a17d4db5c, last modified 2026-09-23 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
Limitations2 values
| EditJumps released model card on Hugging Face No field-specific location recorded Version: Repository revision 44a0ca3e224e16adac0e15e24400d22a17d4db5c, last modified 2026-09-23 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Media type text/markdown Source metadata | EditJumps released model card on Hugging Face No field-specific location recorded Version: Repository revision 44a0ca3e224e16adac0e15e24400d22a17d4db5c, last modified 2026-09-23 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Publication status official_project_source Source metadata | EditJumps released model card on Hugging Face No field-specific location recorded Version: Repository revision 44a0ca3e224e16adac0e15e24400d22a17d4db5c, last modified 2026-09-23 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Retrieved at 2026-11-16T00:06:00Z Source metadata | EditJumps released model card on Hugging Face No field-specific location recorded Version: Repository revision 44a0ca3e224e16adac0e15e24400d22a17d4db5c, last modified 2026-09-23 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Review: artifact sha256 552199e1e3be6d07de661baf23876fcbc354f8cfd33b391d2913291a3a5a8c32 Source metadata | EditJumps released model card on Hugging Face No field-specific location recorded Version: Repository revision 44a0ca3e224e16adac0e15e24400d22a17d4db5c, last modified 2026-09-23 | catalogued No individual claim review recorded Audit detailsField: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Review: method note Independent review pass. Every pinned artifact was re-downloaded from the URL in its source record and its SHA-256 recomputed: all nine match. The arXiv PDF's text layer was regenerated with pdftotext -layout from poppler 26.08.0 and matches the hash notes.md records. npm run batch -- check-values over the whole batch reproduced the extractor's figures exactly (8 in_cell, 164 in_row, 3 locator_not_parsed, no wrong_column and no overclaimed_in_row). Because 164 results sit at in_row only, column assignment was then checked separately by review/column_check.py, written for this pass and using neither extract/tables.py nor extract/check_tables.py. It builds two independent column models: the PDF text layer, where Tables 2 and 4 are assigned by clustering printed field right edges and Table 1 by ordinal position within the row, and the arXiv HTML rendering parsed as a cell grid from the table markup. The two models agree on every cell, and all 172 table-located results match their locator's column and block in both. Four negative controls, two columns exchanged within a row, a row shifted one column left, a Ty1 and HER2 sub-column exchanged, and the Table 1 upper and lower blocks exchanged for one row label, are rejected by both models. Source metadata | EditJumps released model card on Hugging Face No field-specific location recorded Version: Repository revision 44a0ca3e224e16adac0e15e24400d22a17d4db5c, last modified 2026-09-23 | catalogued 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
The model card of the released EditJumps weights, which states the trunk, parameter count, training corpus, objective, optimizer and training budget of the checkpoint.
1 source record and release history
- EditJumps released model card on Hugging Face · Original source · Repository revision 44a0ca3e224e16adac0e15e24400d22a17d4db5c, last modified 2026-09-23
Technical metadata and extraction receipts
Stable ID: editjumps-20261115-source-model-card
- areas
- protein-function; proteins-complexes
- contexts
- research
- access and reuse
- The card declares MIT for the code and the weights. The stock ESM-2 trunk it builds on and the Observed Antibody Space corpus it is trained from carry their own separate terms.
- artifact sha256
- 552199e1e3be6d07de661baf23876fcbc354f8cfd33b391d2913291a3a5a8c32
- artifact url
- https://huggingface.co/VisiumSA/EditJumps/raw/44a0ca3e224e16adac0e15e24400d22a17d4db5c/README.md
- licence
- MIT
- limitations
- Two rows of the 'Model details' table carry the authors' own unresolved HTML comments, 'TODO: confirm head count/names' on the Heads row and 'TODO: confirm step' on the best validation loss row.; Its training cost row and its best validation loss disagree with the paper's Appendix D.
- media type
- text/markdown
- publication status
- official_project_source
- retrieved at
- 2026-11-16T00:06:00Z
- review
- artifact sha256: 552199e1e3be6d07de661baf23876fcbc354f8cfd33b391d2913291a3a5a8c32; method: ai-assisted-source-review; source-hash-verification; method note: Independent review pass. Every pinned artifact was re-downloaded from the URL in its source record and its SHA-256 recomputed: all nine match. The arXiv PDF's text layer was regenerated with pdftotext -layout from poppler 26.08.0 and matches the hash notes.md records. npm run batch -- check-values over the whole batch reproduced the extractor's figures exactly (8 in_cell, 164 in_row, 3 locator_not_parsed, no wrong_column and no overclaimed_in_row). Because 164 results sit at in_row only, column assignment was then checked separately by review/column_check.py, written for this pass and using neither extract/tables.py nor extract/check_tables.py. It builds two independent column models: the PDF text layer, where Tables 2 and 4 are assigned by clustering printed field right edges and Table 1 by ordinal position within the row, and the arXiv HTML rendering parsed as a cell grid from the table markup. The two models agree on every cell, and all 172 table-located results match their locator's column and block in both. Four negative controls, two columns exchanged within a row, a row shifted one column left, a Ty1 and HER2 sub-column exchanged, and the Table 1 upper and lower blocks exchanged for one row label, are rejected by both models.; note: Re-retrieved from the URL in this record and the SHA-256 recomputed; it matches. The repository files were also checked against a tarball of commit d827d47c3bcd29e0f4c15a15f9ac92a6cd859e99, in which they are byte-identical.; retrieval url: https://huggingface.co/VisiumSA/EditJumps/raw/44a0ca3e224e16adac0e15e24400d22a17d4db5c/README.md; reviewed at: 2026-11-16T01:30:00Z; reviewer: claude; reviewer note: Claude (Opus 5) review agent for issue #244, independent of the extraction agent; no human review claimed. All nine pinned artifacts were re-retrieved and their SHA-256 re-computed; no value was reproduced and no model was run.
- source locator
- 'Model details' table; the paragraph below it; 'License'
- url
- https://huggingface.co/VisiumSA/EditJumps
- version
- Repository revision 44a0ca3e224e16adac0e15e24400d22a17d4db5c, last modified 2026-09-23