From 2be29338460a94cafa9013c27c0ec618f481232b Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Tue, 24 Feb 2026 09:26:54 +0000
Subject: [PATCH] skip eval split and HF split detection when eval_steps is
 disabled

---
 studio/backend/core/training/trainer.py  | 41 +++++++++++++-----------
 studio/backend/core/training/training.py |  5 +--
 2 files changed, 24 insertions(+), 22 deletions(-)

diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index ff5b66485a..d433d2575a 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -343,7 +343,8 @@ class UnslothTrainer:
                      custom_format_mapping: dict = None,
                      subset: str = None,
                      train_split: str = "train",
-                     eval_split: str = None) -> Optional[tuple]:
+                     eval_split: str = None,
+                     eval_steps: float = 0.00) -> Optional[tuple]:
         """
         Load and prepare dataset for training.
 
@@ -358,6 +359,7 @@ class UnslothTrainer:
             dataset = None
             eval_dataset = None
             has_separate_eval_source = False  # True if eval comes from a separate HF split
+            eval_enabled = eval_steps is not None and eval_steps > 0
 
             if local_datasets:
                 # Load local datasets
@@ -410,23 +412,26 @@ class UnslothTrainer:
                 print(f"Loaded dataset from Hugging Face: {dataset_source}\n")
 
                 # Resolve eval split from a separate HF split (explicit or auto-detected)
-                if eval_split:
-                    # Explicit eval split provided - load it directly
-                    print(f"Loading explicit eval split: '{eval_split}'\n")
-                    eval_load_kwargs = {"path": dataset_source, "split": eval_split}
-                    if subset:
-                        eval_load_kwargs["name"] = subset
-                    eval_dataset = load_dataset(**eval_load_kwargs)
-                    has_separate_eval_source = True
-                    print(f"Loaded eval split '{eval_split}' with {len(eval_dataset)} rows\n")
-                else:
-                    # Auto-detect eval split from HF (returns a separate dataset, or None)
-                    eval_dataset = self._auto_detect_eval_split_from_hf(
-                        dataset_source=dataset_source,
-                        subset=subset,
-                    )
-                    if eval_dataset is not None:
+                if eval_enabled:
+                    if eval_split:
+                        # Explicit eval split provided - load it directly
+                        print(f"Loading explicit eval split: '{eval_split}'\n")
+                        eval_load_kwargs = {"path": dataset_source, "split": eval_split}
+                        if subset:
+                            eval_load_kwargs["name"] = subset
+                        eval_dataset = load_dataset(**eval_load_kwargs)
                         has_separate_eval_source = True
+                        print(f"Loaded eval split '{eval_split}' with {len(eval_dataset)} rows\n")
+                    else:
+                        # Auto-detect eval split from HF (returns a separate dataset, or None)
+                        eval_dataset = self._auto_detect_eval_split_from_hf(
+                            dataset_source=dataset_source,
+                            subset=subset,
+                        )
+                        if eval_dataset is not None:
+                            has_separate_eval_source = True
+                else:
+                    print("Eval disabled (eval_steps <= 0), skipping eval split detection\n")
 
             if dataset is None:
                 raise ValueError("No dataset provided")
@@ -472,7 +477,7 @@ class UnslothTrainer:
                 )
                 eval_dataset = eval_info["dataset"]
                 print(f"Eval dataset formatted successfully\n")
-            elif not has_separate_eval_source:
+            elif eval_enabled and not has_separate_eval_source:
                 # No separate eval source — split the already-formatted dataset
                 formatted_dataset = dataset_info["dataset"]
                 split_result = self._resolve_eval_split_from_dataset(formatted_dataset)
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index f5d9be63c1..9123d36b39 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -223,6 +223,7 @@ class TrainingBackend:
                 subset=subset,
                 train_split=train_split,
                 eval_split=eval_split,
+                eval_steps=eval_steps,
             )
 
             # Unpack: load_and_format_dataset returns (dataset, eval_dataset)
@@ -232,10 +233,6 @@ class TrainingBackend:
                 dataset = dataset_result
                 eval_dataset = None
 
-            # If user set eval_steps to 0, disable evaluation entirely
-            if eval_steps is not None and float(eval_steps) <= 0:
-                eval_dataset = None
-
             # Track whether eval is enabled for status reporting
             self.eval_enabled = eval_dataset is not None