Update run_eval.py

run_eval.py

@@ -44,7 +44,7 @@ for cfg in CONFIGS:
     adapter_type = cfg.get("adapter_type", "LoRA")
     tasks = cfg["tasks"]
 
-    print(f"\n📦 Loading base model: {base_model_id}")
+    print(f"\nLoading base model: {base_model_id}")
     tokenizer = AutoTokenizer.from_pretrained(base_model_id, use_fast=True)
 
     # Try causal first, fallback to encoder
@@ -110,14 +110,16 @@ with tempfile.TemporaryDirectory() as tmp:
     df_existing = pd.read_parquet(current_path)
     df_combined = pd.concat([df_existing, df_new], ignore_index=True)
 
-    df_combined = (
-        df_combined
-        .sort_values("run_date")
-        .drop_duplicates(subset=["model_id", "task", "metric"], keep="last")
-    )
+    df_combined = df_combined.sort_values("run_date")
+
 
     df_combined["value"] = pd.to_numeric(df_combined["value"], errors="coerce")
 
+    print("Existing rows:", len(df_existing))
+    print("New rows:", len(df_new))
+    print("Combined rows (pre-dedup):", len(df_existing) + len(df_new))
+    print("Final rows (after dedup):", len(df_combined))
+
     out = Path("peft_bench.parquet")
     df_combined.to_parquet(out, index=False)