model card

README.md

@@ -0,0 +1,43 @@
+---
+library_name: transformers
+tags: [mass-spectrometry, proteomics, denoising, msdelta]
+---
+
+# iona-denoise-400m
+
+Per-peak noise classifier for tandem mass spectra: an MSDelta encoder (400m scale,
+396.2M parameters including the head) pretrained on
+`chrisagrams/MSConsensus-100M` for 220,000 steps, then fine-tuned with a token-classification
+head on `chrisagrams/ms-denoise-100k`. One logit per peak; **noise is the positive class
+(label 1)**. Spectra above 512 peaks are out of scope (they were dropped, not truncated,
+in training).
+
+## Results (held-out test split, peptide-disjoint from train)
+
+| metric | value |
+|---|---|
+| AUROC | 0.9438 |
+| AUPRC | 0.9507 |
+| F1 (threshold logit >= 0) | 0.8770 |
+| per-spectrum AUROC (mean) | 0.9487 |
+
+Selected as the best of 3 seeds x the available pretraining checkpoints at this scale by
+**validation** AUPRC (best: 0.9503); the test split was never used
+for selection.
+
+## Training
+
+lr 2e-4 (encoder at 0.5x), 4 epochs, head width 512, effective batch 12 (12 tiles,
+DeepSpeed ZeRO-2), bf16, cosine schedule, best checkpoint by validation AUPRC. Seed
+1. Run `sweep-400m_ck220k_seed1-8850494`.
+
+## Usage
+
+```python
+from transformers import AutoConfig, AutoModelForTokenClassification, AutoProcessor
+repo = "Gaolaboratory/iona-denoise-400m"
+model = AutoModelForTokenClassification.from_pretrained(repo, trust_remote_code=True)
+processor = AutoProcessor.from_pretrained(repo, trust_remote_code=True)
+```
+
+Internal release; not yet reviewed for publication.