Fix LoRA rank to r=4, remove VRAM note, clean up usage sections

README.md

@@ -28,28 +28,22 @@ It corresponds to **model-c** in the [CurveBench paper](https://arxiv.org/abs/26
 
 ## Usage
 
-This is a LoRA adapter — you need the base model alongside it. The recommended way is to serve with **vLLM**, which loads the base model once and applies the adapter on-the-fly.
+### Option 1 — vLLM (recommended for serving)
 
-### 1. Serve with vLLM
-
-```bash
-pip install vllm
-```
+Start the server with the LoRA adapter loaded on top of the base model:
 
 ```bash
 vllm serve google/gemma-3-12b-it \
     --enable-lora \
     --lora-modules grpo-region-tree=AmirMohseni/curvebench-gemma-3-12b \
-    --max-lora-rank 8 \
+    --max-lora-rank 4 \
     --max-model-len 32768 \
     --gpu-memory-utilization 0.90 \
     --dtype bfloat16 \
     --trust-remote-code
 ```
 
-> Requires ~32 GB VRAM (single A100-80GB).
-
-### 2. Query the server (Python)
+Then query it with the OpenAI-compatible API:
 
 ```python
 from openai import OpenAI
@@ -85,21 +79,45 @@ response = client.chat.completions.create(
 print(response.choices[0].message.content)
 ```
 
-### 3. Merge and load with 🤗 Transformers (offline)
+### Option 2 — PEFT + Transformers (offline)
 
-If you prefer to merge the adapter into the base weights without a server:
+Load the base model and apply the LoRA adapter directly:
 
 ```python
 from peft import PeftModel
 from transformers import AutoModelForCausalLM, AutoProcessor
+import torch
 
 base_id = "google/gemma-3-12b-it"
 adapter_id = "AmirMohseni/curvebench-gemma-3-12b"
 
 processor = AutoProcessor.from_pretrained(base_id, trust_remote_code=True)
-model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="auto", device_map="auto", trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(
+    base_id, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
+)
 model = PeftModel.from_pretrained(model, adapter_id)
-model = model.merge_and_unload()  # optional: fuse weights for faster inference
+
+prompt = (
+    "The image shows a set of pairwise non-intersecting closed curves drawn on a plane. "
+    "Each curve creates a boundary between an interior region and its surroundings. "
+    "Output the containment tree of the regions as a list of edges in the format: "
+    "[(parent, child), ...] where 0 is the outermost (unbounded) region."
+)
+
+from PIL import Image
+image = Image.open("curves.png")
+
+inputs = processor(
+    text=processor.apply_chat_template(
+        [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt}]}],
+        add_generation_prompt=True,
+    ),
+    images=[image],
+    return_tensors="pt",
+).to(model.device)
+
+output = model.generate(**inputs, max_new_tokens=512)
+print(processor.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
 ```
 
 ---
@@ -114,7 +132,7 @@ Trained with GRPO using a fork of TRL with multimodal support: [AmirTuring/trl @
 - **Base model:** [google/gemma-3-12b-it](https://huggingface.co/google/gemma-3-12b-it)
 - **Training split:** `total_train` (210 images) from CurveBench-Easy
 - **Reward:** tree isomorphism (0.7) + node count (0.3)
-- **Adapter rank:** 8
+- **LoRA rank (r):** 4 | **LoRA alpha:** 8
 
 ### Framework versions