Update README.md

README.md

@@ -60,7 +60,7 @@ LongWriter-Zero’s effectiveness is demonstrated on two fronts: **WritingBench*
 
 ---
 
-### 🏆 Human-in-the-Loop Win-Rate
+### 🏆 Win-Rate Results
 
  ![image/png](https://cdn-uploads.huggingface.co/production/uploads/63369da91ba5d5ece24118a4/_uWmcqnMFWGLN_iQb1bdx.png)
 
@@ -70,6 +70,46 @@ LongWriter-Zero’s effectiveness is demonstrated on two fronts: **WritingBench*
 
 
 <a name="quick_start"></a>
+## ⚡ Quick Start&nbsp;(HF generate)
+
+```python
+model_name = "THU-KEG/LongWriter-Zero-32B"
+
+model = AutoModelForCausalLM.from_pretrained(
+    model_name,
+    torch_dtype="auto",
+    device_map="auto"
+)
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+prompt = "Write a 500-word story."
+messages = [
+    {"role": "user", "content": prompt}
+]
+text = tokenizer.apply_chat_template(
+    messages,
+    tokenize=False,
+    add_generation_prompt=True
+)
+
+model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
+
+generated_ids = model.generate(
+    **model_inputs,
+    max_new_tokens=2048,
+    stop_strings=["<|user|>", "<|endoftext|>", "</answer>"],
+    tokenizer=tokenizer
+)
+generated_ids = [
+    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
+]
+
+response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+
+print(response)
+```
+*Note: We use a slightly different tokenizer and chat template compared to the original Qwen2.5-32B-Instruct model.*
+
 ## ⚡ Quick Start&nbsp;(SGlang)
 
 The snippet below shows how to format prompts with LongWriter-Zero’s `<think> … </think><answer> … </answer>` protocol and call the model through an SGlang-powered endpoint supporting streaming responses.