From 955cf07271de6095f8ee838a2ce8ab426cdec65c Mon Sep 17 00:00:00 2001
From: David <44320468+gusario@users.noreply.github.com>
Date: Tue, 3 Mar 2026 18:04:21 +0300
Subject: [PATCH] Fix for load_dataset function to restore ability to use
 custom loading of dataset

With update 4.0.0 function load_dataset function creates base DataLoader and don't respect loader field of DatasetMeta. In 3.x.x version it was honored via load_function, so this small change restore ability to use custom load of dataset
---
 swift/dataset/loader.py | 20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/swift/dataset/loader.py b/swift/dataset/loader.py
index ecd98a4dd2..ee803b16b2 100644
--- a/swift/dataset/loader.py
+++ b/swift/dataset/loader.py
@@ -297,16 +297,6 @@ def load_dataset(
         num_proc = None
     train_datasets = []
     val_datasets = []
-    loader = DatasetLoader(
-        num_proc=num_proc,
-        load_from_cache_file=load_from_cache_file,
-        streaming=streaming,
-        hub_token=hub_token,
-        strict=strict,
-        download_mode=download_mode,
-        columns=columns,  # columns_mapping
-        remove_unused_columns=remove_unused_columns,
-    )
 
     use_hf_default = use_hf
     if use_hf_default is None:
@@ -324,6 +314,16 @@ def load_dataset(
                 dataset_syntax.dataset = dataset_meta.hf_dataset_id if use_hf else dataset_meta.ms_dataset_id
         else:
             dataset_meta = dataset_syntax.get_dataset_meta(use_hf)
+        loader = dataset_meta.loader(
+            num_proc=num_proc,
+            load_from_cache_file=load_from_cache_file,
+            streaming=streaming,
+            hub_token=hub_token,
+            strict=strict,
+            download_mode=download_mode,
+            columns=columns,  # columns_mapping
+            remove_unused_columns=remove_unused_columns,
+        )
         train_dataset = loader.load(dataset_syntax, dataset_meta, use_hf=use_hf)
         train_dataset, val_dataset = loader.post_process(
             train_dataset,