Surface per-head CV-AUROC on the taste meters

training/predict.py

@@ -177,13 +177,18 @@ def _ik1(*smiles):
 # load whatever classifier heads exist (sweet/bitter/umami...) + intensity
 _CLASSIFIERS = {}
 _INTENSITY = None
+_TASTE_META = {}  # taste -> {auroc, ...} from taste_models/manifest.json (held-out score)
 if TASTE.exists():
     for p in TASTE.glob("*_rf.joblib"):
         name = p.stem.replace("_rf", "")
         if name == "sweet_intensity":
             _INTENSITY = joblib.load(p)
         else:
             _CLASSIFIERS[name] = joblib.load(p)
+    _tm = TASTE / "manifest.json"
+    if _tm.exists():
+        import json as _json
+        _TASTE_META = _json.loads(_tm.read_text())
 
 # Caution-only toxicity-assay heads (Tox21, public domain). Loaded if trained.
 # INDICATIVE in-vitro signals — never a toxicity determination.
@@ -777,6 +782,10 @@ def predict(smiles: str, include_aroma: bool = False) -> dict:
     }
     for name, clf in sorted(_CLASSIFIERS.items()):
         out[name] = round(float(clf.predict_proba(x)[0, 1]), 3)
+    # per-head held-out CV-AUROC (from taste_models/manifest.json) so the UI can show how
+    # trustworthy each trained taste head is, the same way the aroma descriptors carry theirs
+    out["taste_meta"] = {t: {"auroc": _TASTE_META[t]["auroc"]}
+                         for t in _CLASSIFIERS if t in _TASTE_META and "auroc" in _TASTE_META[t]}
     # Sour trains as a small-data INDICATIVE head, but its boolean stays the rule's
     # call below — keep the model probability separately as sour_predicted.
     if "sour" in out:

training/train_taste.py

@@ -15,6 +15,7 @@
   python train_taste.py
 """
 
+import json
 from pathlib import Path
 
 import joblib
@@ -79,6 +80,7 @@ def featurize(smiles_series):
 def train_classifiers(master):
     X_all, keep = featurize(master["smiles"])
     df = master.iloc[keep].reset_index(drop=True)
+    manifest = {}
     for taste in BASIC:
         y = df[taste]
         mask = y.notna().values
@@ -98,6 +100,8 @@ def train_classifiers(master):
         auc = roc_auc_score(yte, clf.predict_proba(Xte)[:, 1])
         print(f"  {taste:7s} AUROC={auc:.3f}  (pos={pos}, neg={neg})")
         joblib.dump(clf, OUT / f"{taste}_rf.joblib")
+        manifest[taste] = {"auroc": round(float(auc), 3), "n_pos": pos, "n_neg": neg}
+    return manifest
 
 
 def train_intensity(path=Path("sweet_intensity.parquet")):
@@ -113,6 +117,7 @@ def train_intensity(path=Path("sweet_intensity.parquet")):
     r2 = r2_score(yte, reg.predict(Xte))
     print(f"  sweet_intensity R2={r2:.3f}  (n={len(y)})  [small data; treat as indicative]")
     joblib.dump(reg, OUT / "sweet_intensity_rf.joblib")
+    return round(float(r2), 3)
 
 
 def validate_sour_rule(master):
@@ -140,9 +145,12 @@ def validate_sour_rule(master):
 if __name__ == "__main__":
     master = pd.read_parquet("taste_master.parquet")
     print("training taste classifiers:")
-    train_classifiers(master)
+    manifest = train_classifiers(master)
     print("validating the sour rule against labeled data:")
     validate_sour_rule(master)
     print("training sweetness-intensity regressor:")
-    train_intensity()
-    print(f"\nsaved models -> {OUT}/")
+    r2 = train_intensity()
+    if r2 is not None:
+        manifest["sweet_intensity"] = {"r2": r2}
+    (OUT / "manifest.json").write_text(json.dumps(manifest, indent=2))
+    print(f"\nsaved models + manifest.json -> {OUT}/")

training/workbench.html

@@ -42,6 +42,7 @@
   .meter .row{display:flex;justify-content:space-between;align-items:baseline;margin-bottom:5px}
   .meter .name{font-weight:600;font-size:14px;text-transform:capitalize}
   .meter .val{font-family:var(--mono);font-size:13px;color:var(--muted)}
+  .meter .au{font-size:11px;color:var(--muted);opacity:.85}
   .meter .tag{font-family:var(--mono);font-size:10px;padding:2px 7px;border-radius:20px;
     margin-left:8px;vertical-align:middle;letter-spacing:.04em}
   .tag.measured{background:var(--accent-soft);color:var(--accent)}
@@ -232,12 +233,14 @@ <h2>Aroma</h2>
   return r.json();
 }
 
-function meter(name, value, measured){
+function meter(name, value, measured, auroc){
   const pct = Math.round(value*100);
   const tag = measured ? '<span class="tag measured">measured</span>'
                        : '<span class="tag predicted">predicted</span>';
+  // show the head's held-out AUROC on model predictions (not on measured ground truth)
+  const au = (auroc && !measured) ? ` <span class="au">· AUROC ${auroc}</span>` : '';
   return `<div class="meter">
-    <div class="row"><span class="name">${name}${tag}</span><span class="val">${pct}%</span></div>
+    <div class="row"><span class="name">${name}${tag}</span><span class="val">${pct}%${au}</span></div>
     <div class="track"><div class="fill" style="background:${TASTE_COLORS[name]}"
       data-w="${pct}"></div></div></div>`;
 }
@@ -289,7 +292,7 @@ <h2>Aroma</h2>
   if(inDomain){
     let m = '';
     for(const t of MODEL_TASTES){
-      if(typeof p[t] === 'number') m += meter(t, p[t], known.has(t));
+      if(typeof p[t] === 'number') m += meter(t, p[t], known.has(t), (p.taste_meta&&p.taste_meta[t]||{}).auroc);
     }
     if(m) m += '<div class="meters-note">Each is an independent confidence that the molecule carries that taste — they don\'t sum to 100%. A molecule can be both (sweet <i>and</i> bitter) or weakly any.</div>';
     $('meters').innerHTML = m || '<div class="empty">No taste heads trained yet.</div>';