diff --git a/transformers_doc/en/pytorch/training.ipynb b/transformers_doc/en/pytorch/training.ipynb
index 6826183e..c8e5d9b0 100644
--- a/transformers_doc/en/pytorch/training.ipynb
+++ b/transformers_doc/en/pytorch/training.ipynb
@@ -237,6 +237,121 @@
     "trainer.push_to_hub()"
    ]
   },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Native PyTorch Training Loop with Gradient Clipping"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "If you prefer more control over training, you can write a native PyTorch training loop instead of using [Trainer](https://huggingface.co/docs/transformers/main/en/main_classes/trainer#transformers.Trainer).\n",
+    "\n",
+    "One important technique to include is **gradient clipping** (`torch.nn.utils.clip_grad_norm_`), which prevents exploding gradients — a common problem when fine-tuning large models. Exploding gradients cause the loss to become `NaN` or the model weights to grow uncontrollably, destabilizing training.\n",
+    "\n",
+    "First, set up the DataLoader, optimizer, and learning rate scheduler."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "from torch.utils.data import DataLoader\n",
+    "from transformers import AdamW, get_scheduler\n",
+    "from transformers import AutoModelForSequenceClassification, AutoTokenizer\n",
+    "from datasets import load_dataset\n",
+    "\n",
+    "# Load dataset and tokenizer\n",
+    "dataset = load_dataset(\"yelp_review_full\")\n",
+    "tokenizer = AutoTokenizer.from_pretrained(\"google-bert/bert-base-cased\")\n",
+    "\n",
+    "def tokenize(examples):\n",
+    "    return tokenizer(examples[\"text\"], padding=\"max_length\", truncation=True)\n",
+    "\n",
+    "tokenized_dataset = dataset.map(tokenize, batched=True)\n",
+    "\n",
+    "# Use a small subset to keep things fast\n",
+    "small_train = tokenized_dataset[\"train\"].shuffle(seed=42).select(range(1000))\n",
+    "small_eval = tokenized_dataset[\"test\"].shuffle(seed=42).select(range(200))\n",
+    "\n",
+    "# Format datasets for PyTorch\n",
+    "small_train.set_format(\"torch\", columns=[\"input_ids\", \"attention_mask\", \"label\"])\n",
+    "small_eval.set_format(\"torch\", columns=[\"input_ids\", \"attention_mask\", \"label\"])\n",
+    "\n",
+    "train_dataloader = DataLoader(small_train, shuffle=True, batch_size=8)\n",
+    "eval_dataloader = DataLoader(small_eval, batch_size=8)\n",
+    "\n",
+    "# Load model\n",
+    "model = AutoModelForSequenceClassification.from_pretrained(\"google-bert/bert-base-cased\", num_labels=5)\n",
+    "\n",
+    "# Set up optimizer and scheduler\n",
+    "optimizer = AdamW(model.parameters(), lr=5e-5)\n",
+    "num_epochs = 3\n",
+    "num_training_steps = num_epochs * len(train_dataloader)\n",
+    "lr_scheduler = get_scheduler(\n",
+    "    name=\"linear\",\n",
+    "    optimizer=optimizer,\n",
+    "    num_warmup_steps=0,\n",
+    "    num_training_steps=num_training_steps,\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Now write the training loop. Notice the `torch.nn.utils.clip_grad_norm_()` call **after** `loss.backward()` and **before** `optimizer.step()`. This is the correct position — gradients must be computed first before they can be clipped."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from tqdm.auto import tqdm\n",
+    "\n",
+    "device = torch.device(\"cuda\") if torch.cuda.is_available() else torch.device(\"cpu\")\n",
+    "model.to(device)\n",
+    "\n",
+    "progress_bar = tqdm(range(num_training_steps))\n",
+    "\n",
+    "model.train()\n",
+    "for epoch in range(num_epochs):\n",
+    "    for batch in train_dataloader:\n",
+    "        batch = {k: v.to(device) for k, v in batch.items()}\n",
+    "\n",
+    "        # Forward pass\n",
+    "        outputs = model(**batch)\n",
+    "        loss = outputs.loss\n",
+    "\n",
+    "        # Backward pass\n",
+    "        loss.backward()\n",
+    "\n",
+    "        # Gradient clipping — prevents exploding gradients\n",
+    "        # max_norm=1.0 is a safe default for most fine-tuning tasks\n",
+    "        torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)\n",
+    "\n",
+    "        optimizer.step()\n",
+    "        lr_scheduler.step()\n",
+    "        optimizer.zero_grad()\n",
+    "        progress_bar.update(1)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "> [!TIP]\n",
+    "> `max_norm=1.0` is the standard default and works well for most fine-tuning scenarios. If your loss is still unstable, try lowering it to `0.5`. If training is too slow to converge, try raising it to `2.0`."
+   ]
+  },
   {
    "cell_type": "markdown",
    "metadata": {},