| <div class="ai"> | |
| # Evaluation assets | |
| </div> | |
| <div class="ai"> | |
| ## Full-set MLM validation | |
| </div> | |
| <div class="ai"> | |
| The current protocol, `heldout-cluster-representative-mlm-v2`, scores all **12,288 validation proteins**, exactly 4,096 each from UniRef90, MGnify and OMG/IMG. Every protein is evaluated once at a maximum context of 512 tokens. [MLM_VALIDATION_12288.json](MLM_VALIDATION_12288.json) records the protocol and immutable source-shard hashes. The training and validation Parquet files are unchanged. | |
| </div> | |
| <div class="ai"> | |
| Each protein receives a separate PyTorch CPU random generator seeded from mask seed 20260821 and its sequence SHA-256. Proteins longer than 510 residues receive a random contiguous crop, then BOS/EOS framing. Mask each canonical amino-acid position with probability 0.15 and guarantee at least one target when eligible residues exist; replace every selected token with `<mask>`. The contract specifies the seed derivation and random-draw order. | |
| </div> | |
| <div class="ai"> | |
| Crops and masks are constructed before batching, so batch size, evaluation order, GPU count and the global random state do not change them. The default validation batch size is 32 and may be adjusted for throughput or memory. Small numerical differences from floating-point execution remain possible. Compute masked-token NLL within each protein, then average equally over all proteins; each source therefore contributes one third of the overall score. | |
| </div> | |
| <div class="ai"> | |
| All 12,288 sequence digests were recomputed and verified. At context 512, the fixed examples contain **414,065 masked residues**. The contract also records per-source target counts, the evaluated-population digest and a digest of the corrupted inputs and labels for reproducibility. Evaluation receipts identify the v2 protocol, settings, sample counts, per-source losses and checkpoint checksum. | |
| </div> | |
| <div class="ai"> | |
| ## Historical sampled evaluation | |
| </div> | |
| <div class="ai"> | |
| [MLM_VALIDATION_4096.json](MLM_VALIDATION_4096.json) preserves the earlier v1 protocol: 4,096 sampled proteins in 1,024 batches of four, with 1,393 MGnify, 1,378 OMG/IMG and 1,325 UniRef90 proteins and 140,009 masked residues. Its crops and masks depend on the original batch layout. Earlier 32-protein measurements also retain their original settings. Re-evaluate checkpoints under v2 before comparing losses with current results; do not relabel historical scores or reuse v1 cached receipts. | |
| </div> | |
| <div class="ai"> | |
| ## Contact P@L | |
| </div> | |
| <div class="ai"> | |
| [contact-evaluation-v2.tar.gz](contact-evaluation-v2.tar.gz) contains 16 probe-fit chains, four probe-validation chains, all 20,775 evaluation chains and six frozen evaluator source files. [The archive receipt](contact-evaluation-v2.tar.gz.json) provides SHA-256 checksums. The payloads, chain splits and numerical evaluator sources are unchanged; packaging metadata contains dataset attribution without experiment-report links or machine paths. | |
| </div> | |
| <div class="ai"> | |
| PDB structures derive from the 2024-02-28 snapshot and retain their CC0 attribution. Evaluator source retains its accompanying MIT license. The archive includes ATTRIBUTION.md and LICENSE. | |
| </div> | |
| <div class="ai"> | |
| The bundle supports the published contact protocol. P@L averages precision among the top L predicted long-range contacts over all 20,775 chains; use 5,000 chain-bootstrap replicates for its 95% interval. Fit a new probe for each checkpoint with the fixed probe split and procedure. | |
| </div> | |