Readiness
- Readiness
- Production Ready
- Smoke
- passed
- Artifacts
- passed
- Production
- production_ready_api_slurm_balm_paired_masked_lm
- Storage
- <5
- Strategy
- shared python312_slim.sif + polyxpert/palm PyTorch/Transformers envs + Zenodo BALM-paired checkpoint cache under /media/nik/seagate_nik/bio_server/models/balm-paired
- Next
- Production Ready. Optional hardening: add batch mutation-scoring inputs and expose heavy/light mask controls in the UI.
BALM Paired source verified as brineylab/BALM-paper rev aac40d66dea81cca7d95b172f49393001dada0e4; MIT license. Notebook workflow rather than CLI/package/container. Public Zenodo paired training tarball metadata reachable without token (~43.5 MB). Staged size ~28 MB. Slurm job 403 passed scheduler/source/tokenizer checks and confirmed runtime imports missing: torch, transformers, datasets, accelerate. No real mutation-scoring benchmark yet. 2026-07-05 SOP recheck: source/data are small and token-free, but Slurm job 403 only verified scheduler/source/tokenizer checks; torch/transformers/datasets/accelerate runtime and real mutation-scoring smoke remain missing. Kept Smoke Pending. 2026-07-05 SOP batch recheck: source/data are small and token-free, but previous job 403 only checked scheduler/source/tokenizer and found missing torch/transformers/datasets/accelerate; no real mutation-scoring benchmark has passed. Kept Smoke Pending. 2026-07-05 SOP recheck: source/data are small and token-free, but previous job 403 only checked scheduler/source/tokenizer and found missing torch/transformers/datasets/accelerate; no real mutation-scoring benchmark has passed. Kept Smoke Pending. 2026-07-05 SOP batch recheck: source/data are small and token-free, but previous Slurm job 403 only checked scheduler/source/tokenizer and found missing torch/transformers/datasets/accelerate; no real mutation-scoring smoke has passed. Kept Smoke Pending. 2026-07-05 18:19 +08:00 backend re-audit: BALM Paired source/evidence remains staged (~31 MB); previous job 403 only verified scheduler/source/tokenizer and missing torch/transformers/datasets/accelerate. No real mutation-scoring smoke exists. Agents agreed Smoke Pending remains correct. 2026-07-06 20:25 +08:00 direct Slurm smoke passed. Job 470, run /media/nik/seagate_nik/bio_server/runs/balm-paired/smoke_20260706_202500. Downloaded and extracted real BALM-paired pretrained archive from Zenodo record 8253367 (BALM-paired.tar.gz; extracted model BALM-paired_LC-coherence_90-5-5-split_122222; cache size 2.2GB including archive). Loaded RobertaForMaskedLM with vocab_size=25, hidden_size=1024, num_layers=24; ran paired heavy/light masked-residue inference using explicit segment token types. Artifacts: summary.json, results.csv, input_sequence.txt, job.log, slurm.out, slurm.err. Sanity: finite logits/probabilities, top token W probability 0.3265865, actual residue L probability 0.054012 rank 8. Promote to Runner Adapter Pending; API/runner wiring remains next. 2026-07-08T21:05:20+08:00: Production Ready promotion: backend runner API job 621929ea89b3 / Slurm job 568 completed through the web runner API using shared python312_slim.sif with polyxpert_py312_site/palm_py312_site and the staged Zenodo BALM-paired checkpoint. Adapter ran paired heavy/light RobertaForMaskedLM masked-residue inference with explicit light-chain token_type_ids on the 32/32 toy antibody benchmark, produced zero-byte slurm.out/slurm.err, summary.json, input_sequence.txt, model_load_stderr.txt, and results.csv. Scientific sanity: finite probabilities/logits, top token W probability 0.3283219635, actual masked L probability 0.0543432944, actual rank 7, vocab_size 25, 24 layers, hidden_size 1024.