rewirebio.iobenchmarks
Configuration

Evotuned ESM-2 650M, unforced infilling

A stock ESM-2 650M trunk fine-tuned on one antibody family's training partition for 2,000 steps, then infilling entropy-selected masked positions of a template in random order at temperature 1.0, topping up masks over up to eight rounds until the realised substitution count matches the edit budget.

20 evaluations · 20 results

Overview

An explanatory profile has not yet been reviewed for this record.

Consult the linked sources for architecture or protocol details. Missing evidence is not evidence of a missing capability.

Model Capabilities & Hardware Requirements

Deployment facts recorded for Evotuned ESM-2 650M, unforced infilling. Fields the record does not carry are marked not recorded.

ParametersNot recorded
Minimum VRAMNot recorded
Inference LatencyNot recorded
Context WindowNot recorded
AccessNot recorded
LicenseNot recorded

Evaluated Checkpoint: As reimplemented and scored in version 1 of the paper

Evaluations and results

20 evaluations · 20 results. Different protocols are not a single leaderboard.

Sorted by Covariance agreement (residue covariance agreement against a 300-sequence natural homolog ceiling) (higher is better). The best value in each column is highlighted. Decimals are rounded for display; each value links to the printed value and its source.

Tested configurationProtocol and datasetCovariance agreement (residue covariance agreement against a 300-sequence natural homolog ceiling)Jensen shannon divergence (per-position formulation chosen by the EditJumps authors, against a 200-sequence natural holdout)Kullback leibler divergence (per-position amino-acid composition against a 200-sequence natural holdout)Kullback leibler divergence (pooled amino-acid composition against a 200-sequence natural holdout)Log likelihood (profile log-likelihood of the generated sequences under the family alignment profile)Maximum mean discrepancy (k-mer spectrum MMD between the generated set and a 200-sequence natural holdout, unnormalised)Mutual information product agreement (mutual information product agreement against a 300-sequence natural homolog ceiling)Edit count (mean realised edits per generated sequence, 400 generated sequences per family)Entropy delta (positional entropy difference from the natural holdout; the target value is zero)Levenshtein distance (mean pairwise Levenshtein distance within the generated set, pooled over templates)
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: cov. agr., HER2-VH reference holdout, 200 homologs0.93 (best in column)Not reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: cov. agr., Ty1 reference holdout, 200 homologs0.91Not reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: edits, HER2-VH reference holdout, 200 homologsNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported3.83Not reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: edits, Ty1 reference holdout, 200 homologsNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported4.47Not reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: ΔH, HER2-VH reference holdout, 200 homologsNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported0.03Not reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: ΔH, Ty1 reference holdout, 200 homologsNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported0.01Not reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: JS pos., HER2-VH reference holdout, 200 homologsNot reported1.02 (best in column)Not reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: JS pos., Ty1 reference holdout, 200 homologsNot reported1.14Not reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: KL pool., HER2-VH reference holdout, 200 homologsNot reportedNot reportedNot reported0.37 (best in column)Not reportedNot reportedNot reportedNot reportedNot reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: KL pool., Ty1 reference holdout, 200 homologsNot reportedNot reportedNot reported0.54Not reportedNot reportedNot reportedNot reportedNot reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: KL pos., HER2-VH reference holdout, 200 homologsNot reportedNot reported0.05 (best in column)Not reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: KL pos., Ty1 reference holdout, 200 homologsNot reportedNot reported0.06Not reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: MIP agr., HER2-VH reference holdout, 200 homologsNot reportedNot reportedNot reportedNot reportedNot reportedNot reported0.83 (best in column)Not reportedNot reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: MIP agr., Ty1 reference holdout, 200 homologsNot reportedNot reportedNot reportedNot reportedNot reportedNot reported0.81Not reportedNot reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: pairw. Lev., HER2-VH reference holdout, 200 homologsNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported22.8
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: pairw. Lev., Ty1 reference holdout, 200 homologsNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reportedNot reported22.6
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: prof. LL, HER2-VH reference holdout, 200 homologsNot reportedNot reportedNot reportedNot reported-52.9 (best in column)Not reportedNot reportedNot reportedNot reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: prof. LL, Ty1 reference holdout, 200 homologsNot reportedNot reportedNot reportedNot reported-57.5Not reportedNot reportedNot reportedNot reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: MMD, HER2-VH reference holdout, 200 homologsNot reportedNot reportedNot reportedNot reportedNot reported0.99Not reportedNot reportedNot reportedNot reported
Evotuned ESM-2 650M, unforced infillingEvaluationEditJumps section 4.2 frame, 200 references: MMD, Ty1 reference holdout, 200 homologsNot reportedNot reportedNot reportedNot reportedNot reported0.98 (best in column)Not reportedNot reportedNot reportedNot reported
All 20 result rows with coverage, uncertainty and sources
Filter evaluations

Applied filters: All linked evaluations

Exact evaluated configurations and original reported results
Tested configurationProtocol and datasetFindingEvidence and details
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: cov. agr.
Dataset subset: HER2-VH reference holdout, 200 homologs
0.93 Covariance agreement (residue covariance agreement against a 300-sequence natural homolog ceiling)
unitless · higher

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on HER2-VH, cov. agr., 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "her2vh", row "Evotune", column "cov. agr."
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: cov. agr.
Dataset subset: Ty1 reference holdout, 200 homologs
0.91 Covariance agreement (residue covariance agreement against a 300-sequence natural homolog ceiling)
unitless · higher

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on Ty1, cov. agr., 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "ty1", row "Evotune", column "cov. agr."
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: edits
Dataset subset: HER2-VH reference holdout, 200 homologs
3.83 Edit count (mean realised edits per generated sequence, 400 generated sequences per family)
count · unknown

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on HER2-VH, edits, 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "her2vh", row "Evotune", column "edits"
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: edits
Dataset subset: Ty1 reference holdout, 200 homologs
4.47 Edit count (mean realised edits per generated sequence, 400 generated sequences per family)
count · unknown

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on Ty1, edits, 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "ty1", row "Evotune", column "edits"
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: ΔH
Dataset subset: HER2-VH reference holdout, 200 homologs
0.03 Entropy delta (positional entropy difference from the natural holdout; the target value is zero)
unitless · unknown

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on HER2-VH, ΔH, 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "her2vh", row "Evotune", column "ΔH"
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: ΔH
Dataset subset: Ty1 reference holdout, 200 homologs
0.01 Entropy delta (positional entropy difference from the natural holdout; the target value is zero)
unitless · unknown

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on Ty1, ΔH, 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "ty1", row "Evotune", column "ΔH"
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: JS pos.
Dataset subset: HER2-VH reference holdout, 200 homologs
1.02 Jensen shannon divergence (per-position formulation chosen by the EditJumps authors, against a 200-sequence natural holdout)
unitless · lower

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on HER2-VH, JS pos., 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "her2vh", row "Evotune", column "JS pos."
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: JS pos.
Dataset subset: Ty1 reference holdout, 200 homologs
1.14 Jensen shannon divergence (per-position formulation chosen by the EditJumps authors, against a 200-sequence natural holdout)
unitless · lower

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on Ty1, JS pos., 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "ty1", row "Evotune", column "JS pos."
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: KL pool.
Dataset subset: HER2-VH reference holdout, 200 homologs
0.37 Kullback leibler divergence (pooled amino-acid composition against a 200-sequence natural holdout)
unitless · lower

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on HER2-VH, KL pool., 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "her2vh", row "Evotune", column "KL pool."
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: KL pool.
Dataset subset: Ty1 reference holdout, 200 homologs
0.54 Kullback leibler divergence (pooled amino-acid composition against a 200-sequence natural holdout)
unitless · lower

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on Ty1, KL pool., 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "ty1", row "Evotune", column "KL pool."
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: KL pos.
Dataset subset: HER2-VH reference holdout, 200 homologs
0.05 Kullback leibler divergence (per-position amino-acid composition against a 200-sequence natural holdout)
unitless · lower

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on HER2-VH, KL pos., 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "her2vh", row "Evotune", column "KL pos."
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: KL pos.
Dataset subset: Ty1 reference holdout, 200 homologs
0.06 Kullback leibler divergence (per-position amino-acid composition against a 200-sequence natural holdout)
unitless · lower

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on Ty1, KL pos., 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "ty1", row "Evotune", column "KL pos."
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: MIP agr.
Dataset subset: HER2-VH reference holdout, 200 homologs
0.83 Mutual information product agreement (mutual information product agreement against a 300-sequence natural homolog ceiling)
unitless · higher

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on HER2-VH, MIP agr., 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "her2vh", row "Evotune", column "MIP agr."
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: MIP agr.
Dataset subset: Ty1 reference holdout, 200 homologs
0.81 Mutual information product agreement (mutual information product agreement against a 300-sequence natural homolog ceiling)
unitless · higher

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on Ty1, MIP agr., 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "ty1", row "Evotune", column "MIP agr."
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: pairw. Lev.
Dataset subset: HER2-VH reference holdout, 200 homologs
22.8 Levenshtein distance (mean pairwise Levenshtein distance within the generated set, pooled over templates)
count · unknown

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on HER2-VH, pairw. Lev., 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "her2vh", row "Evotune", column "pairw. Lev."
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: pairw. Lev.
Dataset subset: Ty1 reference holdout, 200 homologs
22.6 Levenshtein distance (mean pairwise Levenshtein distance within the generated set, pooled over templates)
count · unknown

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on Ty1, pairw. Lev., 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "ty1", row "Evotune", column "pairw. Lev."
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: prof. LL
Dataset subset: HER2-VH reference holdout, 200 homologs
-52.9 Log likelihood (profile log-likelihood of the generated sequences under the family alignment profile)
unitless · higher

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on HER2-VH, prof. LL, 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "her2vh", row "Evotune", column "prof. LL"
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: prof. LL
Dataset subset: Ty1 reference holdout, 200 homologs
-57.5 Log likelihood (profile log-likelihood of the generated sequences under the family alignment profile)
unitless · higher

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on Ty1, prof. LL, 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "ty1", row "Evotune", column "prof. LL"
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: MMD
Dataset subset: HER2-VH reference holdout, 200 homologs
0.99 Maximum mean discrepancy (k-mer spectrum MMD between the generated set and a 200-sequence natural holdout, unnormalised)
unitless · lower

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on HER2-VH, MMD, 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "her2vh", row "Evotune", column "MMD"
Configuration: Evotuned ESM-2 650M, unforced infillingProtocol: EditJumps section 4.2 frame, 200 references: MMD
Dataset subset: Ty1 reference holdout, 200 homologs
0.98 Maximum mean discrepancy (k-mer spectrum MMD between the generated set and a 200-sequence natural holdout, unnormalised)
unitless · lower

Uncertainty: Not reported by the source: The table prints a single value with no interval.

Coverage: Not reported scored / Not reported eligible

Author-reported evaluation · Source checked
Methods, coverage and source

Evotune on Ty1, MMD, 200 references

One frame for every arm and both families: 20 templates per family and 20 variants per template, so 400 generated sequences, scored against a 200-sequence held-out natural homolog set, with the agreement ceiling computed on 300 real homologs. Sampling runs Euler tau-leaping over 50 steps. Every arm is held to a matched realised edit budget of about 4 to 5 edits per sequence.

Aggregation: Not reported

When Edit Flows are Edit Jumps: replicating Edit Flows and EvoFlows (arXiv v1) · Table 2, family block "ty1", row "Evotune", column "MMD"

Source checking is not independent reproduction. Release 2026-11-21-c8bea25e5abb.

Use this configuration

How it works, versions and access

Additional methodological details, versions and access terms have not yet been extracted into a structured profile for this configuration. Consult the linked sources and evaluation records below.

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.

0 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-11-21-c8bea25e5abb
Property and statementOriginal source and locationReview and provenance

No evidence rows match these filters. Choose another scope or clear the search.

Sources and history

Release 2026-11-21-c8bea25e5abb · Record review: source checked

2 source records and release historyDownload this release (gzip)
Technical metadata and record identifiers

Technical metadata for this record is tracked under stable ID editjumps-20261115-config-evotuning-650m.

Source locator: Section 3 'The evotuning baseline'; Appendix A items 1 to 4; docs/performance.md 'Evotuning ESM-2-650M (A100)'

Related records

How to cite

Evotuned ESM-2 650M, unforced infilling. Rewire biological model benchmark database, release 2026-11-21-c8bea25e5abb. https://benchmarks.rewirebio.io/database/configuration/editjumps-20261115-config-evotuning-650m/

Release 2026-11-21-c8bea25e5abb. Rankings and values change between releases, so keep the release id in the citation. Cite the original publication as well for any value you reuse: each result links its source.

The compilation and annotations made by Rewire in this release are offered under CC BY 4.0; values and text taken from cited sources stay under the terms of those sources and are not covered by this licence. Data licence

Suggest a correction