Initial commit with LLM app

README.md

@@ -1,8 +1,8 @@
 ---
 title: Radiology Report Optimizer
-emoji: 💬
-colorFrom: yellow
-colorTo: purple
+emoji: 🩻
+colorFrom: blue
+colorTo: indigo
 sdk: gradio
 sdk_version: 6.5.1
 app_file: app.py
@@ -10,7 +10,42 @@ pinned: false
 hf_oauth: true
 hf_oauth_scopes:
 - inference-api
-short_description: Radiology Report Refinery
+short_description: Refine radiology reports for clarity and completeness
 ---
 
-An example chatbot using [Gradio](https://gradio.app), [`huggingface_hub`](https://huggingface.co/docs/huggingface_hub/v0.22.2/en/index), and the [Hugging Face Inference API](https://huggingface.co/docs/api-inference/index).
+# Radiology Report Optimizer
+
+A Gradio-powered demo that uses
+[nvidia/Llama-3.1-Nemotron-70B-Instruct-HF](https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Instruct-HF)
+to refine draft radiology reports.
+
+## What it does
+
+Paste a draft radiology report and the model will return an optimized version
+with improvements across six dimensions:
+
+| Dimension | Description |
+|---|---|
+| **Structure** | Standard report sections (Indication, Technique, Comparison, Findings, Impression) |
+| **Clarity** | Precise descriptors and standardized terminology |
+| **Completeness** | Flags missing anatomy, incomplete characterization |
+| **Actionability** | Clear recommendations in the Impression |
+| **Brevity** | Removes redundancy while keeping clinical detail |
+| **Standardized Terminology** | BI-RADS, Lung-RADS, LI-RADS, TI-RADS, Bosniak where applicable |
+
+## Running locally
+
+```bash
+pip install -r requirements.txt
+python app.py
+```
+
+The app will start at `http://localhost:7860`. Log in with your Hugging Face
+account to authenticate with the Inference API.
+
+## Deploying to Hugging Face Spaces
+
+Push this repository to a
+[Hugging Face Space](https://huggingface.co/docs/hub/spaces-overview) with
+**Gradio** as the SDK. The YAML front-matter in this README configures the
+Space automatically.

app.py

@@ -1,67 +1,157 @@
 import gradio as gr
 from huggingface_hub import InferenceClient
 
+MODEL_ID = "nvidia/Llama-3.1-Nemotron-70B-Instruct-HF"
+
+SYSTEM_PROMPT = """\
+You are an expert radiologist AI assistant specializing in optimizing and \
+refining radiology reports. Your task is to take draft radiology reports and \
+improve them for clarity, accuracy, completeness, and adherence to best \
+practices.
+
+When a user provides a radiology report, you should:
+
+1. **Structure** – Ensure the report follows standard radiology report format:
+   - Clinical Indication / History
+   - Technique / Protocol
+   - Comparison (prior studies)
+   - Findings (organized by anatomical system or region)
+   - Impression (concise summary of key findings with numbered items)
+
+2. **Clarity** – Replace vague language with precise descriptors. Use \
+standardized radiology terminology. Avoid hedging language when findings are \
+definitive.
+
+3. **Completeness** – Identify and flag any missing sections, unreported \
+anatomy visible on the study, or findings that lack adequate characterization \
+(size, location, morphology, enhancement pattern, etc.).
+
+4. **Actionability** – Ensure the Impression section includes clear, \
+actionable recommendations when appropriate (e.g., follow-up imaging, \
+correlation with clinical findings, tissue sampling).
+
+5. **Brevity** – Remove redundant phrases and filler words while preserving \
+all clinically relevant information.
+
+6. **Standardized Terminology** – Use ACR BI-RADS, Lung-RADS, LI-RADS, \
+TI-RADS, Bosniak, or other applicable classification systems where relevant.
+
+When providing the optimized report, present the refined version first, then \
+provide a brief summary of the key changes you made. If the user asks \
+follow-up questions, assist with further refinements.
+
+If the user's message is not a radiology report, respond helpfully in the \
+context of radiology reporting best practices.\
+"""
+
+EXAMPLES = [
+    [
+        "Please optimize this CT abdomen report:\n\n"
+        "Clinical: abdominal pain\n\n"
+        "Findings: The liver looks normal. There is a small lesion in the "
+        "right kidney. The pancreas is unremarkable. There is some fluid in "
+        "the pelvis. The appendix is not well seen.\n\n"
+        "Impression: Small kidney lesion. Pelvic fluid. Suggest clinical "
+        "correlation."
+    ],
+    [
+        "Refine this chest X-ray report:\n\n"
+        "History: cough\n\n"
+        "Findings: Heart is normal size. Lungs show some haziness in the "
+        "right lower lobe. No pleural effusion. No pneumothorax. Bones are "
+        "normal.\n\n"
+        "Impression: Right lower lobe opacity, may represent pneumonia."
+    ],
+    [
+        "What are the best practices for structuring a brain MRI report?"
+    ],
+]
+
 
 def respond(
-    message,
+    message: str,
     history: list[dict[str, str]],
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-    hf_token: gr.OAuthToken,
+    max_tokens: int,
+    temperature: float,
+    top_p: float,
+    hf_token: gr.OAuthToken | None = None,
 ):
-    """
-    For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
-    """
-    client = InferenceClient(token=hf_token.token, model="openai/gpt-oss-20b")
-
-    messages = [{"role": "system", "content": system_message}]
+    token = hf_token.token if hf_token else None
+    client = InferenceClient(token=token, model=MODEL_ID)
 
+    messages = [{"role": "system", "content": SYSTEM_PROMPT}]
     messages.extend(history)
-
     messages.append({"role": "user", "content": message})
 
     response = ""
-
-    for message in client.chat_completion(
+    for chunk in client.chat_completion(
         messages,
         max_tokens=max_tokens,
         stream=True,
         temperature=temperature,
         top_p=top_p,
     ):
-        choices = message.choices
-        token = ""
-        if len(choices) and choices[0].delta.content:
-            token = choices[0].delta.content
-
-        response += token
-        yield response
+        choices = chunk.choices
+        if choices and choices[0].delta.content:
+            response += choices[0].delta.content
+            yield response
 
 
-"""
-For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
-"""
 chatbot = gr.ChatInterface(
     respond,
     additional_inputs=[
-        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
-        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
         gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
+            minimum=256, maximum=4096, value=2048, step=64,
+            label="Max new tokens",
+        ),
+        gr.Slider(
+            minimum=0.1, maximum=1.5, value=0.3, step=0.1,
+            label="Temperature",
+        ),
+        gr.Slider(
+            minimum=0.1, maximum=1.0, value=0.9, step=0.05,
             label="Top-p (nucleus sampling)",
         ),
     ],
+    examples=EXAMPLES,
+    title="🩻 Radiology Report Optimizer",
+    description=(
+        "Paste a draft radiology report below to get an optimized version with "
+        "improved structure, clarity, completeness, and adherence to reporting "
+        "standards. Powered by "
+        "**nvidia/Llama-3.1-Nemotron-70B-Instruct-HF**."
+    ),
+    chatbot=gr.Chatbot(
+        height=520,
+        placeholder=(
+            "Paste your radiology report here to get started…\n\n"
+            "You can also ask questions about radiology reporting best practices."
+        ),
+    ),
+    textbox=gr.Textbox(
+        placeholder="Paste your draft radiology report or ask a question…",
+        lines=4,
+    ),
 )
 
-with gr.Blocks() as demo:
+with gr.Blocks(
+    title="Radiology Report Optimizer",
+    theme=gr.themes.Soft(primary_hue="blue", secondary_hue="cyan"),
+) as demo:
     with gr.Sidebar():
         gr.LoginButton()
+        gr.Markdown(
+            "### About\n"
+            "This tool uses **Llama-3.1-Nemotron-70B** to refine "
+            "radiology reports for clarity, structure, and "
+            "completeness.\n\n"
+            "**Tips**\n"
+            "- Paste your full draft report for comprehensive "
+            "optimization\n"
+            "- Ask follow-up questions to refine specific sections\n"
+            "- Request formatting in specific classification systems "
+            "(BI-RADS, Lung-RADS, LI-RADS, etc.)\n"
+        )
     chatbot.render()
 
 

requirements.txt

@@ -0,0 +1,2 @@
+gradio>=4.44.0,<7.0.0
+huggingface_hub>=0.25.0