From 59e079f5e7b79d4986e3c18489b2b15016dd66b1 Mon Sep 17 00:00:00 2001
From: Colin <>
Date: Wed, 13 Aug 2025 16:37:24 +0800
Subject: [PATCH] Update stride test.

---
 wit/doc/train_meaning_dataset.md | 7 ++++++-
 wit/meaning/meaning_dataset.py   | 2 +-
 2 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/wit/doc/train_meaning_dataset.md b/wit/doc/train_meaning_dataset.md
index eddcc43..539e1f7 100644
--- a/wit/doc/train_meaning_dataset.md
+++ b/wit/doc/train_meaning_dataset.md
@@ -17,4 +17,9 @@
 1. key[10] = 1000.0
 2. 每一行数据（像素）表示一个新的token，和前面所有token的关系
 
-![alt text](q@k_seq_47_layer_0.png)
\ No newline at end of file
+![alt text](q@k_seq_47_layer_0.png)
+
+## 在样本的中间插入固定的token
+
+1. 使用stride的方法，在每个token的中间插入一个固定的无用的token
+2. 插入的token用或者不用于计算loss，对精度都没有提升
\ No newline at end of file
diff --git a/wit/meaning/meaning_dataset.py b/wit/meaning/meaning_dataset.py
index 2044cdd..597ddca 100644
--- a/wit/meaning/meaning_dataset.py
+++ b/wit/meaning/meaning_dataset.py
@@ -433,7 +433,7 @@ class MeaningDataset(Dataset):
         val_mask, stride_mask = self.get_seq_mask_tensor(idx_list)
         output["val_mask"] = val_mask
         labels = data.clone()
-        labels[~stride_mask] = self.vocab_size  # set to vocab_size will be masked in label
+        # labels[~stride_mask] = self.vocab_size  # set to vocab_size will be masked in label
         output["labels"] = labels
         output["meaning"] = [self.seq_meaning[i] for i in idx_list]
         return output