The dataset viewer is not available for this split.
Error code: FeaturesError
Exception: ArrowInvalid
Message: Schema at index 1 was different:
weights: string
msa: bool
mode: string
group: string
proposals: int64
observed: int64
assemblies: int64
mean_predicted_probability: double
native_contact_fraction: double
probability_gap: double
gap_low: double
gap_high: double
vs
pdb_id: string
weights: string
msa: bool
tokens: int64
seconds: double
seed: int64
mode: string
sample: int64
group: string
rank: int64
token_i: int64
token_j: int64
probability: double
distance: double
native_observed: bool
native_contact: bool
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 244, in compute_first_rows_from_streaming_response
iterable_dataset = iterable_dataset._resolve_features()
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 4523, in _resolve_features
features = _infer_features_from_batch(self.with_format(None)._head())
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2768, in _head
return next(iter(self.iter(batch_size=n)))
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2972, in iter
for key, pa_table in ex_iterable.iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2483, in _iter_arrow
yield from self.ex_iterable._iter_arrow()
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 591, in _iter_arrow
yield new_key, pa.Table.from_batches(chunks_buffer)
~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^
File "pyarrow/table.pxi", line 5012, in pyarrow.lib.Table.from_batches
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
return check_status(status)
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: Schema at index 1 was different:
weights: string
msa: bool
mode: string
group: string
proposals: int64
observed: int64
assemblies: int64
mean_predicted_probability: double
native_contact_fraction: double
probability_gap: double
gap_low: double
gap_high: double
vs
pdb_id: string
weights: string
msa: bool
tokens: int64
seconds: double
seed: int64
mode: string
sample: int64
group: string
rank: int64
token_i: int64
token_j: int64
probability: double
distance: double
native_observed: bool
native_contact: boolNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Helico exp23: completed masked-contact evaluation
Evaluation of the full-size exp22 model, with released regular/EMA weights, all candidate predictions, official FoldBench scores and immutable input identities. This is a completed evaluation, not a new training run. The model predicts contacts, but FoldBench does not resolve gains on protein–ligand, protein–protein or protein monomer performance; protein–DNA success declines by 3.64 percentage points (nominal 95% interval −7.32 to −0.29). Sparse supplied contacts do not reliably steer structures, and contact-guided sampling does not establish a useful advantage over ordinary batched diffusion.
Official full-input FoldBench
Confidence selects one of four samples using the same one-cycle/200-step inference settings for both weights. The initialization arm uses Protenix v1 weights in Helico; it is not the published default Protenix inference recipe. All 1,522 assemblies and 1,823 category targets remain in the denominators. Interface values below are success rates; monomer values are OpenStructure lDDT with stereochemical checks. Failures/missing inputs count as zero. Intervals bootstrap whole PDB assemblies and retain target weighting.
| Official category | Targets | Initialization | Helico EMA | EMA − initialization (95% CI) |
|---|---|---|---|---|
| Protein–ligand | 558 | 43.9% | 44.4% | +0.5 pp [-2.2, +3.6] |
| Protein–protein | 279 | 58.1% | 57.0% | -1.1 pp [-4.1, +1.8] |
| Antibody–antigen | 172 | 25.0% | 25.0% | +0.0 pp [-5.0, +4.8] |
| Protein–peptide | 51 | 70.6% | 60.8% | -9.8 pp [-23.5, +2.0] |
| Protein–DNA | 330 | 67.6% | 63.9% | -3.6 pp [-7.3, -0.3] |
| Protein–RNA | 70 | 44.3% | 50.0% | +5.7 pp [-2.9, +14.8] |
| Protein monomer | 334 | 0.829 | 0.828 | -0.001 [-0.005, +0.002] |
| RNA monomer | 15 | 0.612 | 0.633 | +0.021 [-0.012, +0.054] |
| DNA monomer | 14 | 0.544 | 0.549 | +0.005 [-0.030, +0.042] |
Coverage: 1482 ready full inputs, 27 resource-cap exclusions and 13 unavailable external alignments; 11856 generated structures. Input/inference and scoring failures are tabulated separately. Oracle best-of-four upper bounds are separate from confidence selection.
Validation findings
- On 384 full complexes, EMA observed-atom lDDT is 0.799787 versus 0.787484 for initialization: paired difference +0.012303 (95% CI +0.008583 to +0.016490). Interface difference +0.008547 (−0.004595 to +0.021932) is unresolved. These fixed-correspondence diagnostics are not official FoldBench scores.
- Macro contact AP with MSA is 0.9066 intraprotein, 0.4847 protein–protein and 0.5345 protein–ligand. MSA removal substantially lowers contact prediction accuracy.
- Up to four supplied missed native contacts recover 3.125%, 3.125% and 6.25% of requested contacts, respectively. No group shows a resolved positive structural-compliance gain over a same-seed unconditioned inference repeat. Oracle diagnostics use validation targets only.
- Progressive unmasking has a small nominal confidence-selected full-structure gain (+0.002564) but no interface, oracle-best or diversity gain. Tree search has no resolved full-structure gain. At matched time, 7-step contact-only sampling is tied; 32 steps lowers interface lDDT by 0.03953 (95% CI −0.08850 to −0.00441). Ordinary sampling remains the primary blind method.
- Two apo/bound reference pairs provide a limited alternative-state diagnostic; contact-guided methods do not outperform ordinary sampling in recovering both states. Diversity alone does not establish valid alternative conformations.
- A mask-support audit shows that the sparsest nonempty training time still reveals about 72 pairs on a fully observed 384-residue protein crop. A subsequent recipe should explicitly cover expected revealed counts 1/4/16/64 with value-independent absorption and verify supplied-constraint compliance before scaling search.
Correction, provenance and limits
The original sampler allowed BF16 autocast to change internal distances during nominally rigid rotation. The correction preserves coordinate precision; model weights did not change. All authoritative structure results use the corrected sampler. Earlier predictions and statistics remain archived for provenance and are superseded. Contact audit-v1 is unaffected. Historical FoldBench predictions diagnosed the defect; corrected validation method selection was frozen before reviewing corrected numerical blind scores. Refer to CURRENT_RESULTS.json for authoritative directories; old files outside this map must not be mixed into current summaries.
No Flock, curated nonbinding, pMHC or TCR training additions were used. Geometric no-contact labels do not establish experimental nonbinding. These measurements do not establish binding specificity or a general conformational ensemble model. Identical inference seeds do not guarantee identical floating-point predictions.
Contents and reproduction
EXPERIMENT.md describes the protocol, results, limitations and released-weight example. configs/ and data/ preserve source/input identities, submission and recovery receipts, full per-target tables and paired summaries. plots/corrected/ contains exportable PNG/SVG figures. raw/<stage>.tar retains every candidate tensor and JSON result, with producer checksums. scored/<stage>.tar.gz contains full predicted mmCIFs and official scorer output/error evidence. Extract each scored archive into a shared directory to resolve the relative prediction paths in the combined CSVs. PyTorch tensor archives require trusted deserialization.
release_manifest.json hashes the analysis snapshot and records remote-verified archive receipts. The unaffected audit archive is additionally pinned in historical release 509631e9036ce59f0875b149545cf0da18765d8c. Legacy local publication receipts are historical references; release_manifest.json and CURRENT_RESULTS.json define this snapshot. The final GitHub notebook records the actual immutable HF revision after its upload verification.
Source and notebook · Evaluation issue · Model weights · Validation inputs
- Downloads last month
- 159


