Configuration: AMPLIFY-350M (HyPLM report) Protocol: NLL, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 1.16 Log loss (Mean negative log-likelihood of masked residues in nats on the report's recent-PDB panel, identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source AMPLIFY-350M on nll-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-nll-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 6, row 'NLL, recent PDB chains', column 'AMPLIFY-350M' (page 12 of 13) Configuration: DPLM-650M (HyPLM report) Protocol: NLL, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 1.8 Log loss (Mean negative log-likelihood of masked residues in nats on the report's recent-PDB panel, identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source DPLM-650M on nll-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-nll-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 6, row 'NLL, recent PDB chains', column 'DPLM-650M' (page 12 of 13) Configuration: E1-600M (HyPLM report) Protocol: NLL, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 1.56 Log loss (Mean negative log-likelihood of masked residues in nats on the report's recent-PDB panel, identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source E1-600M on nll-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-nll-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 6, row 'NLL, recent PDB chains', column 'E1-600M' (page 12 of 13) Configuration: ESM-1b (HyPLM report) Protocol: NLL, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 1.5 Log loss (Mean negative log-likelihood of masked residues in nats on the report's recent-PDB panel, identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ESM-1b on nll-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-nll-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 6, row 'NLL, recent PDB chains', column 'ESM-1b' (page 12 of 13) Configuration: ESM-2 15B (HyPLM report) Protocol: NLL, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 1.16 Log loss (Mean negative log-likelihood of masked residues in nats on the report's recent-PDB panel, identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ESM-2 15B on nll-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-nll-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 3, row 'ESM-2 15B', column 'Recent PDB chains, NLL' (page 6 of 13) Configuration: ESM-2 3B (HyPLM report) Protocol: NLL, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 1.33 Log loss (Mean negative log-likelihood of masked residues in nats on the report's recent-PDB panel, identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ESM-2 3B on nll-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-nll-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 3, row 'ESM-2 3B', column 'Recent PDB chains, NLL' (page 6 of 13) Configuration: ESM-2 650M (HyPLM report) Protocol: NLL, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 1.59 Log loss (Mean negative log-likelihood of masked residues in nats on the report's recent-PDB panel, identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ESM-2 650M on nll-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-nll-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 6, row 'NLL, recent PDB chains', column 'ESM-2 650M' (page 12 of 13) Configuration: ESM C 600M (HyPLM report) Protocol: NLL, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 1.48 Log loss (Mean negative log-likelihood of masked residues in nats on the report's recent-PDB panel, identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ESM C 600M on nll-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-nll-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 6, row 'NLL, recent PDB chains', column 'ESM C 600M' (page 12 of 13) Configuration: ESM C 6B (HyPLM report) Protocol: NLL, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 1.05 Log loss (Mean negative log-likelihood of masked residues in nats on the report's recent-PDB panel, identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ESM C 6B on nll-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-nll-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 3, row 'ESM C 6B', column 'Recent PDB chains, NLL' (page 6 of 13) Configuration: ESM3-open (HyPLM report) Protocol: NLL, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 1.83 Log loss (Mean negative log-likelihood of masked residues in nats on the report's recent-PDB panel, identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ESM3-open on nll-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-nll-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 7, row 'NLL, recent PDB chains', column 'ESM3-open' (page 13 of 13) Configuration: HyPLM-Flash, BF16 evaluation (HyPLM report) Protocol: NLL, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 1.15 Log loss (Mean negative log-likelihood of masked residues in nats on the report's recent-PDB panel, identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source HyPLM-Flash on nll-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-nll-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 6, row 'NLL, recent PDB chains', column 'HyPLM-Flash' (page 12 of 13) Configuration: ProtBERT (HyPLM report) Protocol: NLL, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 1.1 Log loss (Mean negative log-likelihood of masked residues in nats on the report's recent-PDB panel, identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ProtBERT on nll-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-nll-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 7, row 'NLL, recent PDB chains', column 'ProtBERT' (page 13 of 13) Configuration: AMPLIFY-350M (HyPLM report) Protocol: Masked-residue perplexity, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 3.17 Perplexity (Masked-residue perplexity on the recent PDB chains: the exponential of the mean negative log-likelihood of masked residues under the model itself, with identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source AMPLIFY-350M on ppl-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-ppl-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 3, row 'AMPLIFY-350M', column 'Recent PDB chains, PPL' (page 6 of 13) Configuration: DPLM-650M (HyPLM report) Protocol: Masked-residue perplexity, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 6.03 Perplexity (Masked-residue perplexity on the recent PDB chains: the exponential of the mean negative log-likelihood of masked residues under the model itself, with identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source DPLM-650M on ppl-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-ppl-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 3, row 'DPLM-650M', column 'Recent PDB chains, PPL' (page 6 of 13) Configuration: E1-600M (HyPLM report) Protocol: Masked-residue perplexity, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 4.76 Perplexity (Masked-residue perplexity on the recent PDB chains: the exponential of the mean negative log-likelihood of masked residues under the model itself, with identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source E1-600M on ppl-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-ppl-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 3, row 'E1-600M', column 'Recent PDB chains, PPL' (page 6 of 13) Configuration: ESM-2 15B (HyPLM report) Protocol: Masked-residue perplexity, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 3.19 Perplexity (Masked-residue perplexity on the recent PDB chains: the exponential of the mean negative log-likelihood of masked residues under the model itself, with identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ESM-2 15B on ppl-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-ppl-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 3, row 'ESM-2 15B', column 'Recent PDB chains, PPL' (page 6 of 13) Configuration: ESM-2 3B (HyPLM report) Protocol: Masked-residue perplexity, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 3.77 Perplexity (Masked-residue perplexity on the recent PDB chains: the exponential of the mean negative log-likelihood of masked residues under the model itself, with identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ESM-2 3B on ppl-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-ppl-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 3, row 'ESM-2 3B', column 'Recent PDB chains, PPL' (page 6 of 13) Configuration: ESM-2 650M (HyPLM report) Protocol: Masked-residue perplexity, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 4.91 Perplexity (Masked-residue perplexity on the recent PDB chains: the exponential of the mean negative log-likelihood of masked residues under the model itself, with identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ESM-2 650M on ppl-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-ppl-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 3, row 'ESM-2 650M', column 'Recent PDB chains, PPL' (page 6 of 13) Configuration: ESM C 600M (HyPLM report) Protocol: Masked-residue perplexity, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 4.39 Perplexity (Masked-residue perplexity on the recent PDB chains: the exponential of the mean negative log-likelihood of masked residues under the model itself, with identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ESM C 600M on ppl-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-ppl-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 3, row 'ESM C 600M', column 'Recent PDB chains, PPL' (page 6 of 13) Configuration: ESM C 6B (HyPLM report) Protocol: Masked-residue perplexity, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 2.87 Perplexity (Masked-residue perplexity on the recent PDB chains: the exponential of the mean negative log-likelihood of masked residues under the model itself, with identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ESM C 6B on ppl-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-ppl-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 3, row 'ESM C 6B', column 'Recent PDB chains, PPL' (page 6 of 13) Configuration: HyPLM-Flash, BF16 evaluation (HyPLM report) Protocol: Masked-residue perplexity, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 3.15 Perplexity (Masked-residue perplexity on the recent PDB chains: the exponential of the mean negative log-likelihood of masked residues under the model itself, with identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source HyPLM-Flash on ppl-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-ppl-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 3, row 'HyPLM-Flash', column 'Recent PDB chains, PPL' (page 6 of 13) Configuration: ProtBERT (HyPLM report) Protocol: Masked-residue perplexity, recent PDB chains (HyPLM report) Dataset: Recent PDB chains (887, HyPLM report) 3.02 Perplexity (Masked-residue perplexity on the recent PDB chains: the exponential of the mean negative log-likelihood of masked residues under the model itself, with identical masks and crops for every model)unitless · lower
Uncertainty: Not reported by the source: Section 4 says intervals are paired bootstrap intervals; none is printed here.
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons The recent-PDB likelihood panel is described two ways. Section 4.3 and the Table 3 caption give 887 PDB chains released after 2 September 2026; Table 4 gives 887 RCSB PDB entries released 9 September to 7 October 2026. The start of the window and whether the unit is a chain or an entry both differ, so the panel cannot be rebuilt from the report.HyPLM Technical Report (Hyra Team, 2026-10-12) · Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
Appendix A states that the authors could not reproduce xTrimoPGLM's published ProteinGym scores (their runs fall 0.07 to 0.09 below the leaderboard) and that its likelihood is implausibly high, pointing to an incompatibility between their adapter and its masked-language-model head. They withhold its zero-shot and likelihood results and keep its embedding-based results as unaffected, but those results come from the same adapter and that claim was not tested.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Deviations'
Appendix A states that for GB1 two-vs-rest, GB1 three-vs-rest and the cross-protein hydrolase split of FLIP2, validation data come from the training distribution while the test sets are strongly shifted, so probe selection is dominated by noise and numerically equivalent runs of the same model differ by up to 0.2. The authors print these cells but leave them out of every summary statistic.HyPLM Technical Report (Hyra Team, 2026-10-12) · Appendix A 'Statistics'
Methods, coverage and source ProtBERT on ppl-recent-pdb (HyPLM report)
hyplm-flash-20261106-protocol-ppl-recent-pdb
Aggregation: Not reported
HyPLM Technical Report (Hyra Team, 2026-10-12) · Table 3, row 'ProtBERT', column 'Recent PDB chains, PPL' (page 6 of 13)