File size: 6,216 Bytes
21a6776 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 | {
"schema_version": 2,
"protocol": "heldout-cluster-representative-mlm-v2",
"settings": {
"population": "all-validation",
"context_length": 512,
"mask_seed": 20260821
},
"evaluation_sequences": 12288,
"expected_source_counts": {
"uniref90": 4096,
"mgnify": 4096,
"omg_img": 4096
},
"expected_masked_residues": 414065,
"expected_masked_residues_by_source": {
"uniref90": 170557,
"mgnify": 107860,
"omg_img": 135648
},
"reduction": "Mean over all proteins of each protein's mean masked-token negative log-likelihood; equal protein weights give each source one third of the score.",
"selection": {
"population": "Every protein in the three pinned validation shards, exactly once.",
"dataset_selection": "First 4,096 eligible cluster representatives per source in SHA-256 order; their union is excluded from all training source arms.",
"receipt_digest_order": [
"uniref90",
"mgnify",
"omg_img"
],
"within_source_order": "Parquet row order"
},
"per_protein_randomness": {
"generator": "PyTorch 2.13.0 CPU torch.Generator",
"digest_serialization": "bytes.fromhex(protein_sha256).rstrip(b'\\x00')",
"digest_serialization_reason": "v2 reads bytes(store.index[row]['digest']) from a NumPy S32 scalar, which removes trailing zero bytes. Preserve this behavior to reproduce v2 results; the original full SHA-256 remains stored in the Parquet shard.",
"seed_derivation": "int.from_bytes(SHA256(mask_seed.to_bytes(8, 'big') + serialized_digest).digest()[:8], 'big')",
"mask_seed": 20260821,
"draw_order": "If a crop is needed, draw its offset first. Then draw the mask Bernoulli uniforms over the single framed sequence; draw a fallback target index only if no eligible residue was selected."
},
"cropping": {
"maximum_residues": 510,
"maximum_context_tokens": 512,
"offset": "For length > 510, torch.randint(length - 510 + 1, (1,), generator=generator); otherwise zero without an RNG draw.",
"framing": "Prepend BOS and append EOS after cropping."
},
"masking": {
"probability": 0.15,
"eligible": "Canonical amino-acid tokens only; exclude BOS, EOS, padding and noncanonical tokens.",
"selection": "torch.rand((1, framed_length), generator=generator) < 0.15, intersected with eligible positions.",
"minimum_targets": "If any eligible residue exists but none was selected, select one eligible position uniformly with the same generator.",
"corruption": "Replace every selected token with <mask>; labels retain the original token IDs at selected positions and are -100 elsewhere."
},
"batching": {
"batch_size_is_score_setting": false,
"implementation_default_batch_size": 32,
"policy": "Construct crops and masks separately for every protein before batching; group by cropped length and pad within each batch.",
"invariance": "Batch size, evaluation order, GPU count and global RNG state do not select proteins or change their crops and masks. Floating-point execution can introduce small numerical differences."
},
"verification": {
"sequence_digests_recomputed": 12288,
"unique_sequence_digests": 12288,
"evaluated_manifest_sha256": "62a3cf7bde05bf2681b1864f5f45993460f1aa98069029deb47fd70ec36f069f",
"evaluated_manifest_encoding": "Concatenate each protein's digest after the v2 trailing-zero serialization in receipt_digest_order and Parquet row order, then SHA-256; matches validation_mlm.manifest_sha256.",
"canonical_population_sha256": "62968311150cdcb29188c113e7ab83b5b6d757df104fc290e71d8d33b691b48e",
"canonical_population_encoding": "Concatenate each full raw 32-byte SHA-256 digest in the same order, then SHA-256; independent of the mmap scalar serialization.",
"trailing_zero_digest_counts": {
"uniref90": 14,
"mgnify": 15,
"omg_img": 18
},
"examples_sha256": "a8588f0d4620a5823142c7ac6e7f74bdc3a6a62be2a30bc5e4a6dd48b19ced84",
"examples_encoding": "In the same order, concatenate raw protein digest, framed length as uint32 big-endian, corrupted token IDs as int64 little-endian and labels as int64 little-endian; then SHA-256."
},
"validation_pool": {
"manifest_sha256": "fe1ac0657085ab19fe6f56786006e9eb004ca66bc6c5b81dfd8e6bc3dcfda6ff",
"repo_id": "LuminScience/LuminBench-Nano-ESMC",
"revision": "bd38448d50d8f426d7b9bd4410b53159ea001259",
"sources": {
"mgnify": {
"bytes": 578470,
"maximum_sequence_sha256": "0000d08aea015e75fe4a484cf6f66c2d1d70d164bb0aa38389b22960983d53a2",
"minimum_sequence_sha256": "00000003d38602453faede027bc38ed844fb4416607527c22ff799f0d67654f7",
"path": "validation/mgnify/shard-00000.parquet",
"records": 4096,
"residues": 761606,
"sha256": "d9532c1e8490059dfb7178d261b85c31fa7306cd186c0348ee6e3e6ce563fb32"
},
"omg_img": {
"bytes": 724943,
"maximum_sequence_sha256": "00010284808cc5f62cf36bdf8e9f8888f86f9ffd6964ca1a836b9dadef48a742",
"minimum_sequence_sha256": "0000000e7eb16834d3220c994cf1aeb688780a24f5af085705774d0e94d52b7b",
"path": "validation/omg_img/shard-00000.parquet",
"records": 4096,
"residues": 1027103,
"sha256": "a32797838c3338266c62aa61112263cea4515a4a9b977035f88de9cefda08444"
},
"uniref90": {
"bytes": 926829,
"maximum_sequence_sha256": "0003b7ff4ff70e7b21ec32fae89f0792f0d4f86a489346af4877ef8fbad8126b",
"minimum_sequence_sha256": "0000000db96f3a7c2cf445eb4fe4c632a9e047d131200e43131695c51dc2ebdd",
"path": "validation/uniref90/shard-00000.parquet",
"records": 4096,
"residues": 1393942,
"sha256": "f059d9793eb8c7fc929764a37473205d4819652d41baed0e07799bde1055a0fa"
}
},
"total_sequences": 12288
},
"historical_results": {
"previous_contract": "MLM_VALIDATION_4096.json",
"previous_protocol": "heldout-cluster-representative-mlm-v1",
"policy": "Preserve historical sampled results and their original settings. Re-evaluate checkpoints under v2 before comparing with full-population scores; do not relabel old scores or reuse v1 cached evaluation receipts."
}
}
|