Skip to page content
SAVRN
Search Plan a project

Dataset · Other

boltz2_accuracy_architecture_20260919_032813

by Raft Bioworks raftbioworks/boltz2_accuracy_architecture_20260919_032813

Measured small-protein ablations, an evaluation RNG/precision repair for the local Boltz quotient adapter, and a cyclic Fourier triangle-contraction prototype. No new weights were trained. This is not evidence of accurate 100k folding.

Rows—
Configurations—
Size51.3 MB
Licensemit
AccessPublicly accessible
Monthly Downloads—

Dataset Card

By Raft Bioworks, published under mit, revision 5b84958949ab.

Boltz2 accuracy and architecture study

Measured small-protein ablations, an evaluation RNG/precision repair for the local Boltz quotient adapter, and a cyclic Fourier triangle-contraction prototype. No new weights were trained. This is not evidence of accurate 100k folding.

See the report, 38 architecture ideas, and measurements. References are public PDB structures. checkpoint.json identifies the unchanged checkpoint already publicly archived in our preceding audit.

results.tar.gz contains per-case/seed coordinates, input features, status, RNG hashes where recorded, and logs. Failed harness attempts are retained. The old shared_rng field denotes model-entry RNG only; use diffusion_rng_sha256 for the corrected comparison. patch-proof-final uses only the evaluation patch; patch-proof-reference uses both patches. Do not confuse their legacy mode label with the unpatched source.

For reproduction, install the Anthropic Boltz2 kit at f4f62fa6592ae4938d49b1757bea0cfeff9f468e in its separate Python 3.11 / PyTorch 2.12.0+cu130 environment. The launcher references an existing Modal image/volumes in our account; other accounts should build the kit image and mount their own unchanged Boltz2 cache. Set /study to the reproduction scripts and place the chosen original_source/boltz_zeus or fixed_source/boltz_zeus first on PYTHONPATH. The final patched source already includes both patches. point.py uses W&B and a Modal volume commit; adapt those account-specific destinations. Biological runs use 3 recycles, 200 sampling steps, empty MSAs, one sample and seeds 42/43.

All GPU timing measurements are synchronized forward timings rather than download/startup/end-to-end service latency. Native comparisons use complete-assembly alignment, coordinate coverage and pair-weighted C-alpha lDDT, not DockQ. Two seeds on a small selected panel are not an equivalence study. Source code is supplied under its existing licenses; scientific facts and PDB-derived coordinates retain their source provenance.

Details

Repository
raftbioworks/boltz2_accuracy_architecture_20260919_032813
Publisher
Raft Bioworks
Task category
Other
Tags
biology, protein-folding, benchmarking
Size category
Not stated by the source
Languages
Not stated by the source
Revision
5b84958949ab36a4f49b41a4354ccc14cefa2989
Last updated
2026-09-19

Files

94 files, 51.3 MB in total.

Data38 files · 48.5 MB
Documentation5 files · 34.5 KB
Other50 files · 2.8 MB
Repository1 file · 2.5 KB
Every file
FileTypeSizeSHA-256
arithmetic_reservation.jsonData135 B—
arithmetic_status.jsonData108 B—
cases.jsonData8.2 KB—
checkpoint.jsonData223 B—
cyclic-panel_reservation.jsonData145 B—
cyclic-panel_status.jsonData855 B—
diagnostic-c8_reservation.jsonData1.2 KB—
diagnostic-c8_status.jsonData108 B—
diagnostic_reservation.jsonData145 B—
diagnostic_status.jsonData108 B—
fft-micro_reservation.jsonData145 B—
fft-micro_status.jsonData107 B—
fixed-panel_reservation.jsonData146 B—
fixed-panel_status.jsonData854 B—
general-panel_reservation.jsonData135 B—
general-panel_status.jsonData651 B—
manifest.jsonData9.2 KB—
patch-proof-final_reservation.jsonData1.0 KB—
patch-proof-final_status.jsonData215 B—
patch-proof-reference_reservation.jsonData1.1 KB—
patch-proof-reference_status.jsonData215 B—
patch-proof-retry_reservation.jsonData953 B—
patch-proof-retry_status.jsonData214 B—
patch-proof_reservation.jsonData953 B—
patch-proof_status.jsonData215 B—
pilot_reservation.jsonData146 B—
pilot_status.jsonData109 B—
preserve-panel_reservation.jsonData135 B—
preserve-panel_status.jsonData851 B—
reference-panel_reservation.jsonData1.1 KB—
reference-panel_status.jsonData850 B—
repaired-pilot_reservation.jsonData135 B—
repaired-pilot_status.jsonData108 B—
repaired-v2_reservation.jsonData135 B—
repaired-v2_status.jsonData215 B—
results.tar.gzData48.1 MBe72ccd178c5d
scored.csvData112.3 KB—
scored.jsonData291.9 KB—
BOLTZ_LICENSEDocumentation1.1 KB—
FRAMEDIFF_LICENSEDocumentation1.1 KB—
README.mdDocumentation2.2 KB—
docs/boltz_accuracy.mdDocumentation13.0 KB—
docs/boltz_architecture_options.mdDocumentation17.1 KB—
boltz_memory.pyOther4.4 KB—
cyclic_fft.pyOther1.8 KB—
docs/boltz_accuracy_curves.pdfOther57.2 KB—
docs/boltz_accuracy_curves.pngOther535.1 KB6b2dc4ac6b08
docs/boltz_arithmetic_curves.pdfOther41.8 KB—
docs/boltz_arithmetic_curves.pngOther452.5 KBbd1b996617cb
docs/boltz_fft_curves.pdfOther27.8 KB—
docs/boltz_fft_curves.pngOther273.9 KBbe5732a5f5b4
docs/boltz_repaired_curves.pdfOther45.8 KB—
docs/boltz_repaired_curves.pngOther497.7 KB2b538f86d6ba
faithful.pyOther2.5 KB—
fetch_all.pyOther1.3 KB—
fixed_source/boltz_zeus/__init__.pyOther72 B—
fixed_source/boltz_zeus/full_attention.pyOther1.2 KB—
fixed_source/boltz_zeus/full_confidence.pyOther15.9 KB—
fixed_source/boltz_zeus/full_encoders.pyOther6.7 KB—
fixed_source/boltz_zeus/full_model.pyOther8.2 KB—
fixed_source/boltz_zeus/full_trunk.pyOther3.7 KB—
fixed_source/boltz_zeus/quotient.pyOther6.7 KB—
interventions.pyOther6.6 KB—
launch.pyOther3.5 KB—
log_final.pyOther1.5 KB—
micro.pyOther2.1 KB—
model/boltz_memory.pyOther4.4 KB—
model/cyclic_triangle.pyOther2.1 KB—
original_source/boltz_zeus/__init__.pyOther72 B—
original_source/boltz_zeus/full_attention.pyOther1.7 KB—
original_source/boltz_zeus/full_confidence.pyOther15.9 KB—
original_source/boltz_zeus/full_encoders.pyOther6.7 KB—
original_source/boltz_zeus/full_model.pyOther7.8 KB—
original_source/boltz_zeus/full_trunk.pyOther3.4 KB—
original_source/boltz_zeus/metrics.pyOther2.8 KB—
original_source/boltz_zeus/quotient.pyOther6.4 KB—
patches/boltz_quotient_eval.patchOther4.0 KB—
patches/boltz_quotient_reference_attention.patchOther3.9 KB—
plot_arithmetic.pyOther2.2 KB—
point.pyOther7.9 KB—
references/1BRS.cifOther553.6 KB—
references/1UBQ.cifOther103.2 KB—
references/C2_10XK_reference.npzOther2.7 KBbc44e6111c9c
references/C3_1AQ5_reference.npzOther2.4 KBe9114bd57118
references/C4_1C94_reference.npzOther2.3 KBd46aa88a427a
references/C8_1YCY_reference.npzOther6.1 KB8b1a9e5f1c35
references/D2_1A7B_reference.npzOther4.8 KB33e733d7a6e3
references/D3_1BJP_reference.npzOther4.7 KBe2e07f5cb46b
references/T_2BX9_reference.npzOther7.8 KB2bc61833a270
repaired.pyOther4.2 KB—
score_plot.pyOther8.3 KB—
stock_dropout.pyOther763 B—
test_repair.pyOther1.8 KB—
.gitattributesRepository2.5 KB—

License and Download

License
mit
Access
No access gate
Download from Raft Bioworks

Released by Raft Bioworks through its official repository on Hugging Face. Read the license.