rewirebio.iobenchmarks
Source

HyPLM Technical Report (Hyra Team, 2026-10-12)

Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages) · Project source

Original source

What this source contributed

Records in this release that cite this source. A record can cite several sources.

Results

Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.

Model or configurationMetricValueDatasetEvaluation
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation on the FLIP AAV low-vs-high split)0.254 unitlessFLIP AAV, low-vs-high split (FLIP split)AMPLIFY-350M on aav-low-vs-high (HyPLM report)
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation on the FLIP AAV two-vs-many split)0.574 unitlessFLIP AAV, 2-vs-rest split (FLIP split)AMPLIFY-350M on aav-two-vs-many (HyPLM report)
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation on the FLIP2 amylase split partitioned by mutation count)0.298 unitlessAmylase by mutation count (FLIP2, HyPLM report)AMPLIFY-350M on amylase-by-mutation-count (HyPLM report)
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation on the FLIP2 amylase one-to-many split)0.592 unitlessAmylase one-to-many (FLIP2, HyPLM report)AMPLIFY-350M on amylase-one-to-many (HyPLM report)
AMPLIFY-350M (HyPLM report)Mean average precision at 5 (Zero-shot retrieval mean average precision at 5 on DGEB Arch, ranked by cosine similarity between mean-pooled representations)0.057 fractionDGEB Arch retrieval (HyPLM report)AMPLIFY-350M on arch-retrieval-dgeb (HyPLM report)
AMPLIFY-350M (HyPLM report)Precision at l 5 (Long-range contact precision of the top L/5 predictions (sequence separation at least 24) on 82 CASP14 evaluation domains)0.198 fractionCASP14 evaluation domains (82, HyPLM report)AMPLIFY-350M on contacts-casp14 (HyPLM report)
AMPLIFY-350M (HyPLM report)Precision at l 5 (Long-range contact precision of the top L/5 predictions (sequence separation at least 24) on 95 CASP15 evaluation domains)0.244 fractionCASP15 evaluation domains (95, HyPLM report)AMPLIFY-350M on contacts-casp15 (HyPLM report)
AMPLIFY-350M (HyPLM report)Precision at l 5 (Long-range contact precision of the top L/5 predictions (sequence separation at least 24) on the ProteinNet CASP12 test set of 40 proteins)0.210 fractionProteinNet CASP12AMPLIFY-350M on contacts-proteinnet (HyPLM report)
AMPLIFY-350M (HyPLM report)Macro F1 (Macro-averaged F1 of Enzyme Commission classification on DGEB, with the layer and regularisation chosen by 5-fold cross-validation on the training split because this task has no validation split)0.669 fractionDGEB EC classification (HyPLM report)AMPLIFY-350M on ec-classification-dgeb (HyPLM report)
AMPLIFY-350M (HyPLM report)F1 max (Protein-centric Fmax over Enzyme Commission numbers on the DeepFRI EC split)0.861 fractionEC annotation, DeepFRI split (HyPLM report)AMPLIFY-350M on ec-deepfri-split (HyPLM report)
AMPLIFY-350M (HyPLM report)F1 max (Protein-centric Fmax over Enzyme Commission numbers on the SaProt EC split)0.865 fractionEC annotation, SaProt split (HyPLM report)AMPLIFY-350M on ec-saprot-split (HyPLM report)
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation on the TAPE fluorescence regression split)0.660 unitlessFluorescence (TAPE)AMPLIFY-350M on fluorescence (HyPLM report)
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation on the FLIP GB1 low-vs-high split)0.273 unitlessFLIP GB1, low-vs-high split (FLIP split)AMPLIFY-350M on gb1-low-vs-high (HyPLM report)
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation on the FLIP GB1 three-vs-rest split)0.774 unitlessFLIP GB1, 3-vs-rest split (FLIP split)AMPLIFY-350M on gb1-three-vs-rest (HyPLM report)
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation on the FLIP GB1 two-vs-rest split)0.497 unitlessFLIP GB1, 2-vs-rest split (FLIP split)AMPLIFY-350M on gb1-two-vs-rest (HyPLM report)
AMPLIFY-350M (HyPLM report)F1 max (Protein-centric Fmax over Gene Ontology biological-process terms on the DeepFRI GO split as packaged by SaProt)0.467 fractionGO biological process, DeepFRI split via SaProt (HyPLM report)AMPLIFY-350M on go-bp (HyPLM report)
AMPLIFY-350M (HyPLM report)F1 max (Protein-centric Fmax over Gene Ontology cellular-component terms on the DeepFRI GO split as packaged by SaProt)0.524 fractionGO cellular component, DeepFRI split via SaProt (HyPLM report)AMPLIFY-350M on go-cc (HyPLM report)
AMPLIFY-350M (HyPLM report)F1 max (Protein-centric Fmax over Gene Ontology molecular-function terms on the DeepFRI GO split as packaged by SaProt)0.631 fractionGO molecular function, DeepFRI split via SaProt (HyPLM report)AMPLIFY-350M on go-mf (HyPLM report)
AMPLIFY-350M (HyPLM report)Accuracy (Human protein-protein interaction classification accuracy on PEER HumanPPI, with pair features [a, b, a*b, |a-b|])0.811 fractionHumanPPI via PEER (HyPLM report)AMPLIFY-350M on humanppi (HyPLM report)
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation on the FLIP2 cross-protein hydrolase split, training on other hydrolases and testing on P06241)0.212 unitlessHydrolase to P06241 (FLIP2 cross-protein, HyPLM report)AMPLIFY-350M on hydrolase-to-p06241 (HyPLM report)
AMPLIFY-350M (HyPLM report)Accuracy (Ten-class subcellular localisation accuracy on DeepLoc 1.0 via PEER)0.729 fractionDeepLoc ten-class localisation via PEER (HyPLM report)AMPLIFY-350M on localization-10-class (HyPLM report)
AMPLIFY-350M (HyPLM report)Accuracy (Binary localisation accuracy on DeepLoc 1.0 via PEER)0.891 fractionDeepLoc binary localisation via PEER (HyPLM report)AMPLIFY-350M on localization-binary (HyPLM report)
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation on the FLIP Meltome human-cell split)0.665 unitlessFLIP Thermostability, Human-Cell split (FLIP split)AMPLIFY-350M on meltome-human-cell (HyPLM report)
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation on the FLIP Meltome mixed split)0.585 unitlessFLIP Thermostability, Mixed split (FLIP split)AMPLIFY-350M on meltome-mixed (HyPLM report)
AMPLIFY-350M (HyPLM report)Accuracy (Metal-ion binding classification accuracy on the split packaged by SaProt)0.737 fractionMetal-ion binding, SaProt packaging (HyPLM report)AMPLIFY-350M on metal-ion-binding (HyPLM report)
AMPLIFY-350M (HyPLM report)Log loss (Mean negative log-likelihood of masked residues in nats on the report's held-out UniRef panel, identical masks and crops for every model)1.743 unitlessHeld-out UniRef panel (479 proteins, HyPLM report)AMPLIFY-350M on nll-held-out-uniref (HyPLM report)
AMPLIFY-350M (HyPLM report)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)1.155 unitlessRecent PDB chains (887, HyPLM report)AMPLIFY-350M on nll-recent-pdb (HyPLM report)
AMPLIFY-350M (HyPLM report)Perplexity (Masked-residue perplexity on the held-out UniRef panel: the exponential of the mean negative log-likelihood of masked residues under the model itself, with identical masks and crops for every model)5.72 unitlessHeld-out UniRef panel (479 proteins, HyPLM report)AMPLIFY-350M on ppl-held-out-uniref (HyPLM report)
AMPLIFY-350M (HyPLM report)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)3.17 unitlessRecent PDB chains (887, HyPLM report)AMPLIFY-350M on ppl-recent-pdb (HyPLM report)
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation averaged over the 217 substitution assays of ProteinGym v1.3 with the benchmark's official aggregation, from masked-marginal zero-shot scores; per-assay values rounded to three decimals before aggregation)0.363 unitlessProteinGym DMS substitution assays at commit 144fe22 (217 assays)AMPLIFY-350M on proteingym-zero-shot (HyPLM report)
AMPLIFY-350M (HyPLM report)Accuracy (Fold-level classification accuracy on the TAPE remote-homology (SCOP) test set)0.286 fractionFold classification (TAPE)AMPLIFY-350M on remote-homology-fold (HyPLM report)
AMPLIFY-350M (HyPLM report)Accuracy (Three-state (Q3) per-residue secondary-structure accuracy on the CB513 test set)0.797 fractionSecondary structure prediction (TAPE)AMPLIFY-350M on secondary-structure-cb513 (HyPLM report)
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation on the TAPE stability regression split)0.579 unitlessStability (TAPE, HyPLM report)AMPLIFY-350M on stability (HyPLM report)
AMPLIFY-350M (HyPLM report)Spearman (Spearman correlation on the Meltome human-cell split as packaged by SaProt, which the report lists separately from its own FLIP Meltome human-cell row)0.664 unitlessThermostability, Meltome human cell via SaProt (HyPLM report)AMPLIFY-350M on thermostability (HyPLM report)
Ankh-base (HyPLM report)Spearman (Spearman correlation on the FLIP AAV low-vs-high split)0.296 unitlessFLIP AAV, low-vs-high split (FLIP split)Ankh-base on aav-low-vs-high (HyPLM report)
Ankh-base (HyPLM report)Spearman (Spearman correlation on the FLIP AAV two-vs-many split)0.539 unitlessFLIP AAV, 2-vs-rest split (FLIP split)Ankh-base on aav-two-vs-many (HyPLM report)
Ankh-base (HyPLM report)Spearman (Spearman correlation on the FLIP2 amylase split partitioned by mutation count)0.548 unitlessAmylase by mutation count (FLIP2, HyPLM report)Ankh-base on amylase-by-mutation-count (HyPLM report)
Ankh-base (HyPLM report)Spearman (Spearman correlation on the FLIP2 amylase one-to-many split)0.628 unitlessAmylase one-to-many (FLIP2, HyPLM report)Ankh-base on amylase-one-to-many (HyPLM report)
Ankh-base (HyPLM report)Mean average precision at 5 (Zero-shot retrieval mean average precision at 5 on DGEB Arch, ranked by cosine similarity between mean-pooled representations)0.311 fractionDGEB Arch retrieval (HyPLM report)Ankh-base on arch-retrieval-dgeb (HyPLM report)
Ankh-base (HyPLM report)Precision at l 5 (Long-range contact precision of the top L/5 predictions (sequence separation at least 24) on 82 CASP14 evaluation domains)0.467 fractionCASP14 evaluation domains (82, HyPLM report)Ankh-base on contacts-casp14 (HyPLM report)
Ankh-base (HyPLM report)Precision at l 5 (Long-range contact precision of the top L/5 predictions (sequence separation at least 24) on 95 CASP15 evaluation domains)0.523 fractionCASP15 evaluation domains (95, HyPLM report)Ankh-base on contacts-casp15 (HyPLM report)
Ankh-base (HyPLM report)Precision at l 5 (Long-range contact precision of the top L/5 predictions (sequence separation at least 24) on the ProteinNet CASP12 test set of 40 proteins)0.546 fractionProteinNet CASP12Ankh-base on contacts-proteinnet (HyPLM report)
Ankh-base (HyPLM report)Macro F1 (Macro-averaged F1 of Enzyme Commission classification on DGEB, with the layer and regularisation chosen by 5-fold cross-validation on the training split because this task has no validation split)0.680 fractionDGEB EC classification (HyPLM report)Ankh-base on ec-classification-dgeb (HyPLM report)
Ankh-base (HyPLM report)F1 max (Protein-centric Fmax over Enzyme Commission numbers on the DeepFRI EC split)0.870 fractionEC annotation, DeepFRI split (HyPLM report)Ankh-base on ec-deepfri-split (HyPLM report)
Ankh-base (HyPLM report)F1 max (Protein-centric Fmax over Enzyme Commission numbers on the SaProt EC split)0.881 fractionEC annotation, SaProt split (HyPLM report)Ankh-base on ec-saprot-split (HyPLM report)
Ankh-base (HyPLM report)Spearman (Spearman correlation on the TAPE fluorescence regression split)0.642 unitlessFluorescence (TAPE)Ankh-base on fluorescence (HyPLM report)
Ankh-base (HyPLM report)Spearman (Spearman correlation on the FLIP GB1 low-vs-high split)0.386 unitlessFLIP GB1, low-vs-high split (FLIP split)Ankh-base on gb1-low-vs-high (HyPLM report)
Ankh-base (HyPLM report)Spearman (Spearman correlation on the FLIP GB1 three-vs-rest split)0.796 unitlessFLIP GB1, 3-vs-rest split (FLIP split)Ankh-base on gb1-three-vs-rest (HyPLM report)
Ankh-base (HyPLM report)Spearman (Spearman correlation on the FLIP GB1 two-vs-rest split)0.617 unitlessFLIP GB1, 2-vs-rest split (FLIP split)Ankh-base on gb1-two-vs-rest (HyPLM report)
Ankh-base (HyPLM report)F1 max (Protein-centric Fmax over Gene Ontology biological-process terms on the DeepFRI GO split as packaged by SaProt)0.484 fractionGO biological process, DeepFRI split via SaProt (HyPLM report)Ankh-base on go-bp (HyPLM report)

Claims

Statements this source supports, each with where in the source it was checked.

  • HyPLM report: accuracy in bf16: Precision Every accuracy result in the report is evaluated with BF16 arithmetic, for HyPLM-Flash and for all baselines. (Section 3; Section 4)
  • HyPLM report: baseline throughput range: Throughput On an H20, reference implementations of recent 600M-class encoders embed 48,000 to 69,000 residues per second. (Section 1)
  • HyPLM report: block design: Architecture Bidirectional Transformer encoder with RMSNorm pre-normalisation, rotary position embeddings, per-head query-key normalisation and SwiGLU feed-forward layers, trained by masked language modelling. (Section 2.1)
  • HyPLM report: bootstrap: Statistics Confidence intervals use 5,000 bootstrap replicates over assays, proteins, queries or sequences, with identical resamples for paired differences. (Appendix A 'Statistics')
  • HyPLM report: corpus sequences: Training data The corpus holds 2.6 billion sequences (659 billion residues) from UniRef90 and UniRef50, MGnify90, UniParc and JGI IMG, kept at 30 to 8,192 residues. (Section 2.2 'Data')
  • HyPLM report: data ratio e1: Training 5.5 times less training data than E1-600M (4.0 trillion tokens). (Section 2.2 'Training compute')
  • HyPLM report: data ratio esm2: Training 0.73 times the training data of ESM-2 650M. (Section 2.2 'Training compute')
  • HyPLM report: data ratio esmc: Training 8.5 times less training data than ESM C 600M (6.2 trillion tokens). (Section 2.2 'Training compute')
  • HyPLM report: developer: Developer Trained by Hyra 2.0, Tencent Hunyuan's research agent. Authors Haowei Lin, Cai Zhou and Xiangxin Zhou, all affiliated with Tencent Hunyuan. (Abstract; 'Author Information')
  • HyPLM report: figure1 split count: Scope Figure 1 summarises 35 task splits, but Tables 6 and 7 list 32 tasks and Appendix A removes 3 from summaries; the report does not define the 35. (Figure 1 caption and axis; Tables 6 and 7; Appendix A)
  • HyPLM report: figure1b dplm 650m lost: Head to head Figure 1b prints 14 task splits lost by HyPLM-Flash against DPLM-650M, on min-max normalised scores. Figure 1's axis reads 'of 35' but the report never defines its 35 task splits against the 32 tasks of Tables 6 and 7, and the won and lost labels of the ESM-2 650M row sum to 34 while the other seven rows sum to 35, so the denominator is not established. Appendix A also records that three regression splits are left out of Figure 1 altogether. (Figure 1b, row 'DPLM-650M', lost label (page 1))
  • HyPLM report: figure1b dplm 650m won: Head to head Figure 1b prints 21 task splits won by HyPLM-Flash against DPLM-650M, on min-max normalised scores. Figure 1's axis reads 'of 35' but the report never defines its 35 task splits against the 32 tasks of Tables 6 and 7, and the won and lost labels of the ESM-2 650M row sum to 34 while the other seven rows sum to 35, so the denominator is not established. Appendix A also records that three regression splits are left out of Figure 1 altogether. (Figure 1b, row 'DPLM-650M', won label (page 1))
  • HyPLM report: figure1b e1 600m lost: Head to head Figure 1b prints 17 task splits lost by HyPLM-Flash against E1-600M, on min-max normalised scores. Figure 1's axis reads 'of 35' but the report never defines its 35 task splits against the 32 tasks of Tables 6 and 7, and the won and lost labels of the ESM-2 650M row sum to 34 while the other seven rows sum to 35, so the denominator is not established. Appendix A also records that three regression splits are left out of Figure 1 altogether. (Figure 1b, row 'E1-600M', lost label (page 1))
  • HyPLM report: figure1b e1 600m won: Head to head Figure 1b prints 18 task splits won by HyPLM-Flash against E1-600M, on min-max normalised scores. Figure 1's axis reads 'of 35' but the report never defines its 35 task splits against the 32 tasks of Tables 6 and 7, and the won and lost labels of the ESM-2 650M row sum to 34 while the other seven rows sum to 35, so the denominator is not established. Appendix A also records that three regression splits are left out of Figure 1 altogether. (Figure 1b, row 'E1-600M', won label (page 1))
  • HyPLM report: figure1b esm 1b lost: Head to head Figure 1b prints 9 task splits lost by HyPLM-Flash against ESM-1b, on min-max normalised scores. Figure 1's axis reads 'of 35' but the report never defines its 35 task splits against the 32 tasks of Tables 6 and 7, and the won and lost labels of the ESM-2 650M row sum to 34 while the other seven rows sum to 35, so the denominator is not established. Appendix A also records that three regression splits are left out of Figure 1 altogether. (Figure 1b, row 'ESM-1b', lost label (page 1))
  • HyPLM report: figure1b esm 1b won: Head to head Figure 1b prints 26 task splits won by HyPLM-Flash against ESM-1b, on min-max normalised scores. Figure 1's axis reads 'of 35' but the report never defines its 35 task splits against the 32 tasks of Tables 6 and 7, and the won and lost labels of the ESM-2 650M row sum to 34 while the other seven rows sum to 35, so the denominator is not established. Appendix A also records that three regression splits are left out of Figure 1 altogether. (Figure 1b, row 'ESM-1b', won label (page 1))
  • HyPLM report: figure1b esm 2 3b lost: Head to head Figure 1b prints 20 task splits lost by HyPLM-Flash against ESM-2 3B, on min-max normalised scores. Figure 1's axis reads 'of 35' but the report never defines its 35 task splits against the 32 tasks of Tables 6 and 7, and the won and lost labels of the ESM-2 650M row sum to 34 while the other seven rows sum to 35, so the denominator is not established. Appendix A also records that three regression splits are left out of Figure 1 altogether. (Figure 1b, row 'ESM-2 3B', lost label (page 1))
  • HyPLM report: figure1b esm 2 3b won: Head to head Figure 1b prints 15 task splits won by HyPLM-Flash against ESM-2 3B, on min-max normalised scores. Figure 1's axis reads 'of 35' but the report never defines its 35 task splits against the 32 tasks of Tables 6 and 7, and the won and lost labels of the ESM-2 650M row sum to 34 while the other seven rows sum to 35, so the denominator is not established. Appendix A also records that three regression splits are left out of Figure 1 altogether. (Figure 1b, row 'ESM-2 3B', won label (page 1))
  • HyPLM report: figure1b esm 2 650m lost: Head to head Figure 1b prints 14 task splits lost by HyPLM-Flash against ESM-2 650M, on min-max normalised scores. This row's won and lost labels sum to 34 against the axis's 'of 35', while every other row sums to 35; the figure does not say what the remaining split is. (Figure 1b, row 'ESM-2 650M', lost label (page 1))
  • HyPLM report: figure1b esm 2 650m won: Head to head Figure 1b prints 20 task splits won by HyPLM-Flash against ESM-2 650M, on min-max normalised scores. This row's won and lost labels sum to 34 against the axis's 'of 35', while every other row sums to 35; the figure does not say what the remaining split is. (Figure 1b, row 'ESM-2 650M', won label (page 1))
  • HyPLM report: figure1b esm c 300m lost: Head to head Figure 1b prints 12 task splits lost by HyPLM-Flash against ESM C 300M, on min-max normalised scores. Figure 1's axis reads 'of 35' but the report never defines its 35 task splits against the 32 tasks of Tables 6 and 7, and the won and lost labels of the ESM-2 650M row sum to 34 while the other seven rows sum to 35, so the denominator is not established. Appendix A also records that three regression splits are left out of Figure 1 altogether. (Figure 1b, row 'ESM C 300M', lost label (page 1))
  • HyPLM report: figure1b esm c 300m won: Head to head Figure 1b prints 23 task splits won by HyPLM-Flash against ESM C 300M, on min-max normalised scores. Figure 1's axis reads 'of 35' but the report never defines its 35 task splits against the 32 tasks of Tables 6 and 7, and the won and lost labels of the ESM-2 650M row sum to 34 while the other seven rows sum to 35, so the denominator is not established. Appendix A also records that three regression splits are left out of Figure 1 altogether. (Figure 1b, row 'ESM C 300M', won label (page 1))
  • HyPLM report: figure1b esm c 600m lost: Head to head Figure 1b prints 14 task splits lost by HyPLM-Flash against ESM C 600M, on min-max normalised scores. Figure 1's axis reads 'of 35' but the report never defines its 35 task splits against the 32 tasks of Tables 6 and 7, and the won and lost labels of the ESM-2 650M row sum to 34 while the other seven rows sum to 35, so the denominator is not established. Appendix A also records that three regression splits are left out of Figure 1 altogether. (Figure 1b, row 'ESM C 600M', lost label (page 1))
  • HyPLM report: figure1b esm c 600m won: Head to head Figure 1b prints 21 task splits won by HyPLM-Flash against ESM C 600M, on min-max normalised scores. Figure 1's axis reads 'of 35' but the report never defines its 35 task splits against the 32 tasks of Tables 6 and 7, and the won and lost labels of the ESM-2 650M row sum to 34 while the other seven rows sum to 35, so the denominator is not established. Appendix A also records that three regression splits are left out of Figure 1 altogether. (Figure 1b, row 'ESM C 600M', won label (page 1))
  • HyPLM report: figure1b esm3 open lost: Head to head Figure 1b prints 14 task splits lost by HyPLM-Flash against ESM3-open, on min-max normalised scores. Figure 1's axis reads 'of 35' but the report never defines its 35 task splits against the 32 tasks of Tables 6 and 7, and the won and lost labels of the ESM-2 650M row sum to 34 while the other seven rows sum to 35, so the denominator is not established. Appendix A also records that three regression splits are left out of Figure 1 altogether. (Figure 1b, row 'ESM3-open', lost label (page 1))

Models, benchmarks and other records

Models (5)

Configurations (22)

Benchmarks (1)

Protocols (42)

Datasets (22)

Evaluations (484)

Sources (5)

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.

24 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-11-21-adc15377fab2
Property and statementOriginal source and locationReview and provenance
Archive note
The PDF is not committed: it is refetchable at the pinned commit and extract/parse.py refuses any other bytes. Its gzipped text layer is artifacts/HyPLM_Technical_Report.txt.gz.
Source metadata
HyPLM Technical Report (Hyra Team, 2026-10-12)

Original source ↗

No field-specific location recorded

Version: Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages)
Retrieved: 2026-11-06T11:08:21Z

catalogued

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: attributes.archive_note

Source artifact SHA-256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label, metric label and cell count asserted.

Inspected artifact

Archive sha256
e051bbea5e0c0ddb2f61f11cfb0fb9247c51b031cbd5568db70e25e720b3067f
Source metadata
HyPLM Technical Report (Hyra Team, 2026-10-12)

Original source ↗

No field-specific location recorded

Version: Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages)
Retrieved: 2026-11-06T11:08:21Z

catalogued

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: attributes.archive_sha256

Source artifact SHA-256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label, metric label and cell count asserted.

Inspected artifact

Archive transformation receipt
gzip -n9 of the pdftotext -layout output of the pinned PDF (poppler pdftotext, macOS).
Source metadata
HyPLM Technical Report (Hyra Team, 2026-10-12)

Original source ↗

No field-specific location recorded

Version: Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages)
Retrieved: 2026-11-06T11:08:21Z

catalogued

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: attributes.archive_transformation_receipt

Source artifact SHA-256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label, metric label and cell count asserted.

Inspected artifact

Artifact SHA-256
d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710
Source metadata
HyPLM Technical Report (Hyra Team, 2026-10-12)

Original source ↗

No field-specific location recorded

Version: Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages)
Retrieved: 2026-11-06T11:08:21Z

catalogued

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: attributes.artifact_sha256

Source artifact SHA-256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label, metric label and cell count asserted.

Inspected artifact

Artifact url
https://raw.githubusercontent.com/Tencent-Hunyuan/Hyra-results/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hyplm_flash/HyPLM_Technical_Report.pdf
Source metadata
HyPLM Technical Report (Hyra Team, 2026-10-12)

Original source ↗

No field-specific location recorded

Version: Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages)
Retrieved: 2026-11-06T11:08:21Z

catalogued

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: attributes.artifact_url

Source artifact SHA-256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label, metric label and cell count asserted.

Inspected artifact

Evidence concerns
3 values
  • applies to
    • hyplm-flash-20261106-data-nll-recent-pdb
    artifact sha256
    d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710
    message
    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.
    review method
    ai-assisted-source-review
    reviewed at
    2026-11-06T11:08:21Z
    source id
    hyplm-flash-20261106-source-report
    source locator
    Section 4.3; Table 3 caption; Table 4 row 'PDB time split'
  • applies to
    • hyplm-flash-20261106-config-xtrimopglm-1b
    artifact sha256
    d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710
    message
    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.
    review method
    ai-assisted-source-review
    reviewed at
    2026-11-06T11:08:21Z
    source id
    hyplm-flash-20261106-source-report
    source locator
    Appendix A 'Deviations'
  • applies to
    • hyplm-flash-20261106-protocol-gb1-two-vs-rest
    • hyplm-flash-20261106-protocol-gb1-three-vs-rest
    • hyplm-flash-20261106-protocol-hydrolase-to-p06241
    artifact sha256
    d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710
    message
    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.
    review method
    ai-assisted-source-review
    reviewed at
    2026-11-06T11:08:21Z
    source id
    hyplm-flash-20261106-source-report
    source locator
    Appendix A 'Statistics'
Source metadata
HyPLM Technical Report (Hyra Team, 2026-10-12)

Original source ↗

No field-specific location recorded

Version: Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages)
Retrieved: 2026-11-06T11:08:21Z

catalogued

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: attributes.evidence_concerns

Source artifact SHA-256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label, metric label and cell count asserted.

Inspected artifact

Extraction method
pdftotext -layout text layer, parsed by extract/parse.py with every row label, metric label and cell count asserted.
Source metadata
HyPLM Technical Report (Hyra Team, 2026-10-12)

Original source ↗

No field-specific location recorded

Version: Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages)
Retrieved: 2026-11-06T11:08:21Z

catalogued

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: attributes.extraction_method

Source artifact SHA-256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label, metric label and cell count asserted.

Inspected artifact

Licence
Apache-2.0 (repository LICENSE)
Source metadata
HyPLM Technical Report (Hyra Team, 2026-10-12)

Original source ↗

No field-specific location recorded

Version: Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages)
Retrieved: 2026-11-06T11:08:21Z

catalogued

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: attributes.licence

Source artifact SHA-256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label, metric label and cell count asserted.

Inspected artifact

Limitations
6 values
  • Author-released report, not peer reviewed. All three authors are at Tencent Hunyuan, which released the model, and every baseline was run by the same team through its own adapters; no result here is independent of the developer.
  • All accuracy results were computed with BF16 arithmetic. The released default is the INT8 engine, for which the report prints no per-task value; it states only that INT8 keeps zero-shot fitness, structure, function annotation and likelihood within 0.01 of the BF16 results.
  • Section 4 states that confidence intervals are paired bootstrap intervals, but the report prints intervals only for four ProteinGym paired differences. Every table cell is printed without its interval.
  • Tables 6 and 7 print every task for 13 of the 29 models only. The other 16 checkpoints appear at most in Figure 3a, Table 3, Table 5 or the prose, and 8 of the 28 public checkpoints are never named with a value, so the report's list of 28 cannot be reconstructed from it.
  • Table 5's rows without a dagger are published ProteinGym leaderboard values at commit 144fe22, quoted rather than run by the authors; they are not recorded from this source.
  • Throughput was measured on one NVIDIA H20 with the INT8 engine for HyPLM-Flash and with reference implementations for the baselines, each with the fastest attention kernel its implementation supports. Only HyPLM-Flash's own figure and the baselines' range are printed.
Source metadata
HyPLM Technical Report (Hyra Team, 2026-10-12)

Original source ↗

No field-specific location recorded

Version: Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages)
Retrieved: 2026-11-06T11:08:21Z

catalogued

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: attributes.limitations

Source artifact SHA-256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label, metric label and cell count asserted.

Inspected artifact

Media type
application/pdf
Source metadata
HyPLM Technical Report (Hyra Team, 2026-10-12)

Original source ↗

No field-specific location recorded

Version: Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages)
Retrieved: 2026-11-06T11:08:21Z

catalogued

No individual claim review recorded

Source has a recorded evidence concern. Consult its source page before using the claim.

Audit details

Field: attributes.media_type

Source artifact SHA-256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710

Hash scope: pdftotext -layout text layer, parsed by extract/parse.py with every row label, metric label and cell count asserted.

Inspected artifact

How to read the evidence

Sources and history

Release 2026-11-21-adc15377fab2 · Record review: source checked

Technical report for HyPLM-Flash, with its architecture, training, INT8 inference engine and a 32-task evaluation against 28 public protein language model checkpoints.

Release history

No supporting source is linked yet.

Read original source

Download this release (gzip)
Technical metadata and extraction receipts

Stable ID: hyplm-flash-20261106-source-report

areas
protein-function; proteins-complexes
contexts
research
archive note
The PDF is not committed: it is refetchable at the pinned commit and extract/parse.py refuses any other bytes. Its gzipped text layer is artifacts/HyPLM_Technical_Report.txt.gz.
archive sha256
e051bbea5e0c0ddb2f61f11cfb0fb9247c51b031cbd5568db70e25e720b3067f
archive transformation receipt
gzip -n9 of the pdftotext -layout output of the pinned PDF (poppler pdftotext, macOS).
artifact sha256
d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710
artifact url
https://raw.githubusercontent.com/Tencent-Hunyuan/Hyra-results/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hyplm_flash/HyPLM_Technical_Report.pdf
evidence concerns
applies to: hyplm-flash-20261106-data-nll-recent-pdb; artifact sha256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710; message: 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.; review method: ai-assisted-source-review; reviewed at: 2026-11-06T11:08:21Z; source id: hyplm-flash-20261106-source-report; source locator: Section 4.3; Table 3 caption; Table 4 row 'PDB time split'; applies to: hyplm-flash-20261106-config-xtrimopglm-1b; artifact sha256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710; message: 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.; review method: ai-assisted-source-review; reviewed at: 2026-11-06T11:08:21Z; source id: hyplm-flash-20261106-source-report; source locator: Appendix A 'Deviations'; applies to: hyplm-flash-20261106-protocol-gb1-two-vs-rest; hyplm-flash-20261106-protocol-gb1-three-vs-rest; hyplm-flash-20261106-protocol-hydrolase-to-p06241; artifact sha256: d81d88a415a91c2b00d8e97513a90c03ccffd857ec75033fa7de77227dea1710; message: 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.; review method: ai-assisted-source-review; reviewed at: 2026-11-06T11:08:21Z; source id: hyplm-flash-20261106-source-report; source locator: Appendix A 'Statistics'
extraction method
pdftotext -layout text layer, parsed by extract/parse.py with every row label, metric label and cell count asserted.
licence
Apache-2.0 (repository LICENSE)
limitations
Author-released report, not peer reviewed. All three authors are at Tencent Hunyuan, which released the model, and every baseline was run by the same team through its own adapters; no result here is independent of the developer.; All accuracy results were computed with BF16 arithmetic. The released default is the INT8 engine, for which the report prints no per-task value; it states only that INT8 keeps zero-shot fitness, structure, function annotation and likelihood within 0.01 of the BF16 results.; Section 4 states that confidence intervals are paired bootstrap intervals, but the report prints intervals only for four ProteinGym paired differences. Every table cell is printed without its interval.; Tables 6 and 7 print every task for 13 of the 29 models only. The other 16 checkpoints appear at most in Figure 3a, Table 3, Table 5 or the prose, and 8 of the 28 public checkpoints are never named with a value, so the report's list of 28 cannot be reconstructed from it.; Table 5's rows without a dagger are published ProteinGym leaderboard values at commit 144fe22, quoted rather than run by the authors; they are not recorded from this source.; Throughput was measured on one NVIDIA H20 with the INT8 engine for HyPLM-Flash and with reference implementations for the baselines, each with the fastest attention kernel its implementation supports. Only HyPLM-Flash's own figure and the baselines' range are printed.
media type
application/pdf
publication status
official_project_source
retrieved at
2026-11-06T11:08:21Z
source type
vendor technical report
url
https://github.com/Tencent-Hunyuan/Hyra-results/blob/0dcc533cc9f8f76c90ecbfe944df816b6672328a/AI4Science/hyplm_flash/HyPLM_Technical_Report.pdf
version
Dated 2026-10-12 on page 1; retrieved at commit 0dcc533cc9f8f76c90ecbfe944df816b6672328a (13 pages)
version note
The report is dated 2026-10-12 on page 1. The PDF's own CreationDate and ModDate are both 2026-10-29T23:28:47+08 (pdfTeX-1.40.19), so these bytes were built 17 days after the printed date. The printed date is what version records. No earlier PDF exists at the pinned commit, so whether the rebuild changed any content was not checked.
Related records

Suggest a correction