The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
git_commit: string
source_sha256: string
python: string
packages: struct<torch: string, transformers: string, wandb: string, datasets: string, numpy: string, huggingf (... 16 chars omitted)
child 0, torch: string
child 1, transformers: string
child 2, wandb: string
child 3, datasets: string
child 4, numpy: string
child 5, huggingface-hub: string
torch: string
cuda: string
gpu: string
new_proteins: int64
new_residues: int64
seconds: double
residues_per_second: double
peak_gpu_bytes: int64
complete: bool
cache_config: struct<attention: string, batching: string, biohub_revision: string, dense_storage: string, esmc_rev (... 297 chars omitted)
child 0, attention: string
child 1, batching: string
child 2, biohub_revision: string
child 3, dense_storage: string
child 4, esmc_revision: string
child 5, extractor_sha256: string
child 6, fold_revision: string
child 7, format: int64
child 8, layer: int64
child 9, max_length: int64
child 10, mixture_accumulation: string
child 11, precision: string
child 12, sae_revision: string
child 13, sae_values: string
child 14, special_tokens: string
child 15, te_enabled: bool
child 16, te_nvrtc_disabled: bool
child 17, transformer_engine: string
proteins: int64
dataset: struct<revision: string, private: bool, files: int64, url: string>
child 0, revision: string
child 1, private: bool
child 2, files: int64
child 3, url: string
runtime_environment: struct<git_commit: string, source_sha256: string,
...
d 18, nvidia-cudnn-cu13: string
child 19, nvidia-cusolver: string
child 20, nvidia-cusparselt-cu13: string
child 21, nvidia-cuda-nvrtc-cu12: string
child 22, nvidia-nccl-cu13: string
child 23, nvidia-cuda-cupti: string
child 24, nvidia-cusparse: string
child 25, nvidia-cuda-runtime-cu12: string
child 26, xformers: string
child 4, torch: string
child 5, cuda: string
child 6, gpu: string
child 7, nvidia_smi: string
child 8, nvcc: string
child 9, extension_build_nvcc: string
child 10, mmseqs: string
shards: list<item: struct<name: string, bytes: int64, sha256: string>>
child 0, item: struct<name: string, bytes: int64, sha256: string>
child 0, name: string
child 1, bytes: int64
child 2, sha256: string
config: struct<attention: string, batching: string, biohub_revision: string, dense_storage: string, esmc_rev (... 297 chars omitted)
child 0, attention: string
child 1, batching: string
child 2, biohub_revision: string
child 3, dense_storage: string
child 4, esmc_revision: string
child 5, extractor_sha256: string
child 6, fold_revision: string
child 7, format: int64
child 8, layer: int64
child 9, max_length: int64
child 10, mixture_accumulation: string
child 11, precision: string
child 12, sae_revision: string
child 13, sae_values: string
child 14, special_tokens: string
child 15, te_enabled: bool
child 16, te_nvrtc_disabled: bool
child 17, transformer_engine: string
to
{'config': {'attention': Value('string'), 'batching': Value('string'), 'biohub_revision': Value('string'), 'dense_storage': Value('string'), 'esmc_revision': Value('string'), 'extractor_sha256': Value('string'), 'fold_revision': Value('string'), 'format': Value('int64'), 'layer': Value('int64'), 'max_length': Value('int64'), 'mixture_accumulation': Value('string'), 'precision': Value('string'), 'sae_revision': Value('string'), 'sae_values': Value('string'), 'special_tokens': Value('string'), 'te_enabled': Value('bool'), 'te_nvrtc_disabled': Value('bool'), 'transformer_engine': Value('string')}, 'proteins': Value('int64'), 'dataset': {'revision': Value('string'), 'private': Value('bool'), 'files': Value('int64'), 'url': Value('string')}, 'runtime_environment': {'git_commit': Value('string'), 'source_sha256': Value('string'), 'python': Value('string'), 'packages': {'torch': Value('string'), 'transformers': Value('string'), 'wandb': Value('string'), 'datasets': Value('string'), 'numpy': Value('string'), 'huggingface-hub': Value('string'), 'nvidia-nvjitlink': Value('string'), 'nvidia-curand': Value('string'), 'nvidia-cublas-cu12': Value('string'), 'nvidia-cufft': Value('string'), 'nvidia-cublas': Value('string'), 'nvidia-nvshmem-cu13': Value('string'), 'nvidia-cuda-nvrtc': Value('string'), 'nvidia-cufile': Value('string'), 'nvidia-cuda-runtime': Value('string'), 'nvidia-cudnn-frontend': Value('string'), 'nvidia-ml-py': Value('string'), 'nvidia-nvtx': Value('string'), 'nvidia-cudnn-cu13': Value('string'), 'nvidia-cusolver': Value('string'), 'nvidia-cusparselt-cu13': Value('string'), 'nvidia-cuda-nvrtc-cu12': Value('string'), 'nvidia-nccl-cu13': Value('string'), 'nvidia-cuda-cupti': Value('string'), 'nvidia-cusparse': Value('string'), 'nvidia-cuda-runtime-cu12': Value('string'), 'xformers': Value('string')}, 'torch': Value('string'), 'cuda': Value('string'), 'gpu': Value('string'), 'nvidia_smi': Value('string'), 'nvcc': Value('string'), 'extension_build_nvcc': Value('string'), 'mmseqs': Value('string')}, 'shards': List({'name': Value('string'), 'bytes': Value('int64'), 'sha256': Value('string')})}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
git_commit: string
source_sha256: string
python: string
packages: struct<torch: string, transformers: string, wandb: string, datasets: string, numpy: string, huggingf (... 16 chars omitted)
child 0, torch: string
child 1, transformers: string
child 2, wandb: string
child 3, datasets: string
child 4, numpy: string
child 5, huggingface-hub: string
torch: string
cuda: string
gpu: string
new_proteins: int64
new_residues: int64
seconds: double
residues_per_second: double
peak_gpu_bytes: int64
complete: bool
cache_config: struct<attention: string, batching: string, biohub_revision: string, dense_storage: string, esmc_rev (... 297 chars omitted)
child 0, attention: string
child 1, batching: string
child 2, biohub_revision: string
child 3, dense_storage: string
child 4, esmc_revision: string
child 5, extractor_sha256: string
child 6, fold_revision: string
child 7, format: int64
child 8, layer: int64
child 9, max_length: int64
child 10, mixture_accumulation: string
child 11, precision: string
child 12, sae_revision: string
child 13, sae_values: string
child 14, special_tokens: string
child 15, te_enabled: bool
child 16, te_nvrtc_disabled: bool
child 17, transformer_engine: string
proteins: int64
dataset: struct<revision: string, private: bool, files: int64, url: string>
child 0, revision: string
child 1, private: bool
child 2, files: int64
child 3, url: string
runtime_environment: struct<git_commit: string, source_sha256: string,
...
d 18, nvidia-cudnn-cu13: string
child 19, nvidia-cusolver: string
child 20, nvidia-cusparselt-cu13: string
child 21, nvidia-cuda-nvrtc-cu12: string
child 22, nvidia-nccl-cu13: string
child 23, nvidia-cuda-cupti: string
child 24, nvidia-cusparse: string
child 25, nvidia-cuda-runtime-cu12: string
child 26, xformers: string
child 4, torch: string
child 5, cuda: string
child 6, gpu: string
child 7, nvidia_smi: string
child 8, nvcc: string
child 9, extension_build_nvcc: string
child 10, mmseqs: string
shards: list<item: struct<name: string, bytes: int64, sha256: string>>
child 0, item: struct<name: string, bytes: int64, sha256: string>
child 0, name: string
child 1, bytes: int64
child 2, sha256: string
config: struct<attention: string, batching: string, biohub_revision: string, dense_storage: string, esmc_rev (... 297 chars omitted)
child 0, attention: string
child 1, batching: string
child 2, biohub_revision: string
child 3, dense_storage: string
child 4, esmc_revision: string
child 5, extractor_sha256: string
child 6, fold_revision: string
child 7, format: int64
child 8, layer: int64
child 9, max_length: int64
child 10, mixture_accumulation: string
child 11, precision: string
child 12, sae_revision: string
child 13, sae_values: string
child 14, special_tokens: string
child 15, te_enabled: bool
child 16, te_nvrtc_disabled: bool
child 17, transformer_engine: string
to
{'config': {'attention': Value('string'), 'batching': Value('string'), 'biohub_revision': Value('string'), 'dense_storage': Value('string'), 'esmc_revision': Value('string'), 'extractor_sha256': Value('string'), 'fold_revision': Value('string'), 'format': Value('int64'), 'layer': Value('int64'), 'max_length': Value('int64'), 'mixture_accumulation': Value('string'), 'precision': Value('string'), 'sae_revision': Value('string'), 'sae_values': Value('string'), 'special_tokens': Value('string'), 'te_enabled': Value('bool'), 'te_nvrtc_disabled': Value('bool'), 'transformer_engine': Value('string')}, 'proteins': Value('int64'), 'dataset': {'revision': Value('string'), 'private': Value('bool'), 'files': Value('int64'), 'url': Value('string')}, 'runtime_environment': {'git_commit': Value('string'), 'source_sha256': Value('string'), 'python': Value('string'), 'packages': {'torch': Value('string'), 'transformers': Value('string'), 'wandb': Value('string'), 'datasets': Value('string'), 'numpy': Value('string'), 'huggingface-hub': Value('string'), 'nvidia-nvjitlink': Value('string'), 'nvidia-curand': Value('string'), 'nvidia-cublas-cu12': Value('string'), 'nvidia-cufft': Value('string'), 'nvidia-cublas': Value('string'), 'nvidia-nvshmem-cu13': Value('string'), 'nvidia-cuda-nvrtc': Value('string'), 'nvidia-cufile': Value('string'), 'nvidia-cuda-runtime': Value('string'), 'nvidia-cudnn-frontend': Value('string'), 'nvidia-ml-py': Value('string'), 'nvidia-nvtx': Value('string'), 'nvidia-cudnn-cu13': Value('string'), 'nvidia-cusolver': Value('string'), 'nvidia-cusparselt-cu13': Value('string'), 'nvidia-cuda-nvrtc-cu12': Value('string'), 'nvidia-nccl-cu13': Value('string'), 'nvidia-cuda-cupti': Value('string'), 'nvidia-cusparse': Value('string'), 'nvidia-cuda-runtime-cu12': Value('string'), 'xformers': Value('string')}, 'torch': Value('string'), 'cuda': Value('string'), 'gpu': Value('string'), 'nvidia_smi': Value('string'), 'nvcc': Value('string'), 'extension_build_nvcc': Value('string'), 'mmseqs': Value('string')}, 'shards': List({'name': Value('string'), 'bytes': Value('int64'), 'sha256': Value('string')})}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Frozen ESMC-6B representations
The complete full-corpus-v1 cache contains 1,037,086 protein records in 1,095 content-addressed shards, totaling 1,173,424,539,904 feature bytes. It covers all 1,032,662 proteins required by the reconstructed AtlasV2 baseline, plus 4,424 previously cached proteins. SQLite indexes and manifests add a small amount of storage.
Pinned cache receipts identify the four immutable snapshots: the reused cache and extraction partitions 0, 1, and 2. Download from each receipt's exact revision. Verification checks every referenced shard and index against its published SHA256 and byte size. Complete snapshots are immutable; interrupted downloads can reuse objects already present locally.
| Field | Residue or protein shape | Stored dtype |
|---|---|---|
layer60 |
(l, 2560) |
Lossless BF16 bits in uint16 |
structural |
(l, 256) |
Lossless BF16 bits in uint16 |
sae_indices |
(l, 64) |
uint16 |
sae_values |
(l, 64) |
FP32 |
sae_pooled |
(16384,) |
FP32 |
The frozen extraction uses the pinned official Biohub ESMC-6B implementation, the layer-60 SAE, and pretrained ESMFold2 residue-mixture weights. Each manifest records exact source revisions and preprocessing identity. Extraction uses BF16, with no FP8 cache substitution. The structural tensor is the residue mixture before quadratic pair expansion.
Sequences are limited to a 1,024-residue prefix during extraction, with special tokens removed explicitly. Protein lookup uses the SHA256 of the original full sequence. The index records each field's shape, dtype, byte offset, byte count, and checksum. BF16 bit patterns must be reinterpreted as BF16 rather than numerically cast from integers. Sparse SAE features avoid a persistent dense residue-by-16,384 array.
Download shards to disk before training and use read-only SQLite and memory-mapped binary arrays. Each loader worker should open its own handles. The AtlasV2 download_snapshot interface verifies a pinned manifest, downloads shared objects once, verifies their checksums, and publishes the local index last. Multiple snapshots can be merged through SQLite and hard links without copying feature bytes. Complete coverage spans more than 1,024 shards, so full-cache readers must allow sufficient file descriptors.
AtlasV2 baseline training uses the first 512 cached residues. Those representations include context from the longer cached prefix. Cache membership is not a training split: representations for the E. coli benchmark are present for evaluation, and benchmark rows are never training examples. Split membership, homology exclusions, and supervision remain separately versioned.
The repository was moved from lhallee/AtlasV2-embeddings to public lhallee/ESMC_embeddings under explicit owner authorization after a repeated private-storage failure. The move receipt records preserved history. Historical pinned revisions remain accessible. Model and training-data repositories were not made public by this change.
Sources and license restrictions
Protein sources include the released Synthyra PPI snapshot, Synthyra DDI snapshot, and E. coli benchmark. Predicted DDI evidence and experimental PPI evidence are distinct supervision sources.
Retain the Biohub license, Biohub third-party notices, and FlashPPI license. FlashPPI includes CC BY-NC 4.0 restrictions. Public availability does not broaden any upstream license permissions.
- Downloads last month
- 49