feat: auto-write eval/metrics.json on every eval run — gradio_app loads live metrics

eval/evaluator.py

@@ -4,6 +4,7 @@ Main evaluation orchestrator for bias detection framework.
 This module coordinates the evaluation process and provides the main interface
 for running evaluations.
 """
+import json
 from datetime import datetime
 from pathlib import Path
 from typing import List, Optional
@@ -139,7 +140,26 @@ class BiasEvaluationOrchestrator:
             csv_filename = f"f1_report_{timestamp}.csv"
             csv_path = self.results_writer.write_csv_report(csv_data, csv_filename)
             print(f"Report saved to: {csv_path}")
-            
+
+            # Write metrics.json for gradio_app.py to load at startup
+            lang_map = {
+                Language.ENGLISH: "en", Language.SWAHILI: "sw",
+                Language.FRENCH: "fr", Language.GIKUYU: "ki",
+            }
+            metrics_out = {}
+            for r in results:
+                code = lang_map.get(r.language, r.language.value)
+                m = r.overall_metrics
+                metrics_out[code] = {
+                    "f1": round(m.f1_score, 3),
+                    "precision": round(m.precision, 3),
+                    "recall": round(m.recall, 3),
+                    "samples": m.true_positives + m.false_positives + m.false_negatives + m.true_negatives,
+                }
+            metrics_path = Path(__file__).resolve().parent / "metrics.json"
+            metrics_path.write_text(json.dumps(metrics_out, indent=2))
+            print(f"Metrics saved to: {metrics_path}")
+
         except Exception as e:
             print(f"Warning: Failed to save results: {e}")
 

eval/metrics.json

@@ -0,0 +1,26 @@
+{
+  "en": {
+    "f1": 0.786,
+    "precision": 1.0,
+    "recall": 0.647,
+    "samples": 66
+  },
+  "sw": {
+    "f1": 0.816,
+    "precision": 0.733,
+    "recall": 0.92,
+    "samples": 64723
+  },
+  "fr": {
+    "f1": 0.542,
+    "precision": 1.0,
+    "recall": 0.371,
+    "samples": 50
+  },
+  "ki": {
+    "f1": 0.352,
+    "precision": 0.926,
+    "recall": 0.217,
+    "samples": 11848
+  }
+}

gradio_app.py

@@ -36,15 +36,26 @@ LANGS = {
     "Gikuyu":   ("ki", Language.GIKUYU),
 }
 
+# Load live metrics from eval/metrics.json if available (written by eval/evaluator.py)
+_METRICS_PATH = Path(__file__).parent / "eval" / "metrics.json"
+_LIVE_METRICS: dict = {}
+try:
+    _LIVE_METRICS = json.loads(_METRICS_PATH.read_text())
+except Exception:
+    pass
+
+def _m(code: str, field: str, fallback):
+    return _LIVE_METRICS.get(code, {}).get(field, fallback)
+
 # Per-model metrics: model_key -> lang_code -> metrics dict
 MODEL_METRICS = {
     "rules": {
         "label": "Rules-based (lexicon)",
         "description": "Deterministic lexicon rules across all 4 languages. High precision, no GPU needed.",
-        "en": dict(f1=0.885, precision=1.000, recall=0.794, tier="Pre-Bronze", samples=66),
-        "sw": dict(f1=0.821, precision=0.741, recall=0.919, tier="Gold (sample count)", samples=64_723),
-        "fr": dict(f1=0.793, precision=1.000, recall=0.657, tier="Pre-Bronze", samples=50),
-        "ki": dict(f1=0.368, precision=0.916, recall=0.231, tier="Bronze (sample count)", samples=11_848),
+        "en": dict(f1=_m("en","f1",0.885), precision=_m("en","precision",1.000), recall=_m("en","recall",0.794), tier="Pre-Bronze", samples=_m("en","samples",66)),
+        "sw": dict(f1=_m("sw","f1",0.821), precision=_m("sw","precision",0.741), recall=_m("sw","recall",0.919), tier="Gold (sample count)", samples=_m("sw","samples",64_723)),
+        "fr": dict(f1=_m("fr","f1",0.793), precision=_m("fr","precision",1.000), recall=_m("fr","recall",0.657), tier="Pre-Bronze", samples=_m("fr","samples",50)),
+        "ki": dict(f1=_m("ki","f1",0.368), precision=_m("ki","precision",0.916), recall=_m("ki","recall",0.231), tier="Bronze (sample count)", samples=_m("ki","samples",11_848)),
     },
     "ml_classifier": {
         "label": "sw-bias-classifier-v2 (ML)",