From f56efdb4abd6bac88c5d8f69a4649ce4495dbad8 Mon Sep 17 00:00:00 2001
From: VED <146507396+ved1beta@users.noreply.github.com>
Date: Mon, 16 Mar 2026 07:41:23 +0530
Subject: [PATCH] fix: high eval loss w/ sample packing  (#3478) [skip ci]

* check if eval_sp

* radable condition
---
 src/axolotl/utils/data/sft.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/src/axolotl/utils/data/sft.py b/src/axolotl/utils/data/sft.py
index 69cbfb871..e008b542b 100644
--- a/src/axolotl/utils/data/sft.py
+++ b/src/axolotl/utils/data/sft.py
@@ -348,7 +348,9 @@ def _load_raw_datasets(
             dataset = handle_long_seq_in_dataset(dataset, cfg.eval_sequence_len, cfg)
         else:
             dataset = handle_long_seq_in_dataset(dataset, cfg.sequence_len, cfg)
-        if cfg.sample_packing:
+        if (split == "train" and cfg.sample_packing) or (
+            split == "test" and cfg.eval_sample_packing
+        ):
             dataset, _ = process_datasets_for_packing(cfg, dataset, None)
 
         # Deduplicate before saving so the saved dataset is already de-duplicated