Update README.md

README.md

@@ -1,11 +1,12 @@
 ---
 license: apache-2.0
-library_name: transformers
+library_name: peft
 pipeline_tag: text-generation
 base_model: Qwen/Qwen3.6-27B
 tags:
 - unsloth
 - lora
+- peft
 - finetune
 - cybersecurity
 - security
@@ -13,15 +14,48 @@ tags:
 
 # Changeway-Qwen3.6-27B-V1
 
-This model is a fine-tuned version of [Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B), specifically optimized and enhanced for the **Cybersecurity** domain. It was efficiently trained using [Unsloth](https://github.com/unslothai/unsloth) and LoRA (Low-Rank Adaptation).
+⚠️ **Note: This repository contains the LoRA adapter weights only.** It is not a standalone model. You must load it alongside the base model [Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B).
+
+This LoRA adapter is fine-tuned to enhance the model's capabilities specifically in the **Cybersecurity** domain. It was efficiently trained using [Unsloth](https://github.com/unslothai/unsloth).
 
 ## 🛡️ Domain Focus: Cybersecurity
-This model has been fine-tuned on specialized cybersecurity datasets to improve its capabilities in:
+This fine-tuned adapter improves the base model's performance in:
 * Threat intelligence analysis
 * Log analysis and incident response
 * General cybersecurity knowledge retrieval
 
+## 💻 How to Merge and Save (Unsloth)
+
+You can easily download this LoRA adapter, merge it with the base model, and save it as a complete 16-bit model using Unsloth. 
+
+Make sure you have Unsloth installed, then run the following Python script:
+
+```python
+from unsloth import FastLanguageModel
+
+# 1. Point model_name directly to this Hugging Face repository!
+# Unsloth will automatically read the config and load the base model together with the LoRA.
+LORA_DIR = "CTCT-CT2/Changeway-Qwen3.6-27B-LoRA-V1" 
+
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name = LORA_DIR, 
+    max_seq_length = 8192,
+    dtype = None,
+    load_in_4bit = False, # Note: It is best to disable 4bit when merging, load in 16bit mode
+    device_map = "auto",  # [!] Let it automatically take over or force allocation
+)
+
+# 2. Merge the LoRA into the Base model and save as a new full model
+MERGED_DIR = "./qwen-27b-cybersec-merged"
+print(f"Merging and saving to {MERGED_DIR} ...")
+
+model.save_pretrained_merged(MERGED_DIR, tokenizer, save_method="merged_16bit")
+print("Merge completed successfully!")
+
+```
+
 ## 🚀 Training Details
+
 * **Base Model:** [Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B)
 * **Training Framework:** [Unsloth](https://github.com/unslothai/unsloth)
-* **Method:** LoRA
+* **Method:** LoRA (Low-Rank Adaptation)