fix vLLM instructions for adapter

README.md

@@ -93,6 +93,10 @@ incident response chain. Keep an analyst in the loop.
 
 ## Getting Started
 
+This repository holds the **LoRA adapter** (~640 MB), not merged weights. The base model
+`unsloth/Qwen3.6-27B` is downloaded separately and the adapter is applied on top, so both loaders below
+pull roughly 55 GB on first run.
+
 ### Transformers + Unsloth
 
 ```python
@@ -119,21 +123,28 @@ print(ask("Analyze the following CVE description and map it to the appropriate C
 
 ### vLLM
 
-Merge the adapter into the base, then serve text-only:
+Serve the base text-only and attach the adapter as a LoRA module:
 
 ```bash
-vllm serve BeyondMemory/BeyondCTI-27B \
+vllm serve unsloth/Qwen3.6-27B \
+  --enable-lora \
+  --lora-modules beyondcti=BeyondMemory/BeyondCTI-27B \
+  --max-lora-rank 32 \
   --language-model-only \
   --reasoning-parser qwen3 \
   --max-model-len 16384
 ```
 
+Then call it with `model="beyondcti"`. If you would rather serve a single merged checkpoint, merge
+locally first with `PeftModel.from_pretrained(base, adapter).merge_and_unload()` and point vLLM at the
+output directory.
+
 ```python
 from openai import OpenAI
 client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
 
 r = client.chat.completions.create(
-    model="BeyondMemory/BeyondCTI-27B", temperature=0,
+    model="beyondcti", temperature=0,
     messages=[
         {"role": "system", "content": "You are a cybersecurity expert specializing in cyberthreat intelligence."},
         {"role": "user", "content": "Analyze the following CVE description and calculate the CVSS v3.1 "