From 271d1c9ed1f86fa1fce0c6e409590a8184a9f43a Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=D0=9C=D0=B0=D0=B7=D1=83=D1=80=20=D0=93=D1=80=D0=B5=D1=82?=
 =?UTF-8?q?=D0=B0=20=D0=95=D0=B2=D0=B3=D0=B5=D0=BD=D1=8C=D0=B5=D0=B2=D0=BD?=
 =?UTF-8?q?=D0=B0?= <gemazur_1@edu.hse.ru>
Date: Thu, 27 Mar 2025 13:13:27 +0300
Subject: [PATCH] supermega

---
 .ipynb_checkpoints/ULTRAMegaOB-checkpoint.py | 12 ++++++------
 ULTRAMegaOB.py                               | 12 ++++++------
 2 files changed, 12 insertions(+), 12 deletions(-)

diff --git a/.ipynb_checkpoints/ULTRAMegaOB-checkpoint.py b/.ipynb_checkpoints/ULTRAMegaOB-checkpoint.py
index e6d7bd7..1d00d46 100644
--- a/.ipynb_checkpoints/ULTRAMegaOB-checkpoint.py
+++ b/.ipynb_checkpoints/ULTRAMegaOB-checkpoint.py
@@ -141,10 +141,10 @@ def compute_metrics(p):
                 "support": attack_report[attack_type]["support"]
             }
     metrics = {
-    'accuracy': safety_report["accuracy"],
-    'f1': safety_report["weighted avg"]["f1-score"],
-    'unsafe_recall': safety_report["unsafe"]["recall"],  # Без eval_ префикса
-    'safe_precision': safety_report["safe"]["precision"],
+    'eval_accuracy': safety_report["accuracy"],
+    'eval_f1': safety_report["weighted avg"]["f1-score"],
+    'eval_unsafe_recall': safety_report["unsafe"]["recall"],  # Добавляем eval_ префикс
+    'eval_safe_precision': safety_report["safe"]["precision"],
     }
     
     # Добавляем метрики для атак (только если есть unsafe примеры)
@@ -553,7 +553,7 @@ def train_model():
             logging_steps=100,
             save_total_limit=2,
             # load_best_model_at_end=True,
-            metric_for_best_model="unsafe_recall",
+            # metric_for_best_model="unsafe_recall",
             # greater_is_better=True,
             fp16=True,  # Принудительное использование mixed precision
             fp16_full_eval=True,
@@ -561,7 +561,7 @@ def train_model():
             report_to="none",
             seed=Config.SEED, 
             max_grad_norm=1.0,
-            # metric_for_best_model="eval_unsafe_recall",
+            metric_for_best_model="eval_unsafe_recall", 
             greater_is_better=True,
             load_best_model_at_end=True,
         )
diff --git a/ULTRAMegaOB.py b/ULTRAMegaOB.py
index e6d7bd7..1d00d46 100644
--- a/ULTRAMegaOB.py
+++ b/ULTRAMegaOB.py
@@ -141,10 +141,10 @@ def compute_metrics(p):
                 "support": attack_report[attack_type]["support"]
             }
     metrics = {
-    'accuracy': safety_report["accuracy"],
-    'f1': safety_report["weighted avg"]["f1-score"],
-    'unsafe_recall': safety_report["unsafe"]["recall"],  # Без eval_ префикса
-    'safe_precision': safety_report["safe"]["precision"],
+    'eval_accuracy': safety_report["accuracy"],
+    'eval_f1': safety_report["weighted avg"]["f1-score"],
+    'eval_unsafe_recall': safety_report["unsafe"]["recall"],  # Добавляем eval_ префикс
+    'eval_safe_precision': safety_report["safe"]["precision"],
     }
     
     # Добавляем метрики для атак (только если есть unsafe примеры)
@@ -553,7 +553,7 @@ def train_model():
             logging_steps=100,
             save_total_limit=2,
             # load_best_model_at_end=True,
-            metric_for_best_model="unsafe_recall",
+            # metric_for_best_model="unsafe_recall",
             # greater_is_better=True,
             fp16=True,  # Принудительное использование mixed precision
             fp16_full_eval=True,
@@ -561,7 +561,7 @@ def train_model():
             report_to="none",
             seed=Config.SEED, 
             max_grad_norm=1.0,
-            # metric_for_best_model="eval_unsafe_recall",
+            metric_for_best_model="eval_unsafe_recall", 
             greater_is_better=True,
             load_best_model_at_end=True,
         )
-- 
GitLab