chore: update max new tokes

app.py

@@ -29,7 +29,8 @@ if not os.path.exists(file_path):
 llm = LlamaCppLLM(
     model_path=file_path,
     n_gpu_layers=-1,
-    n_ctx=1024 * 128,
+    # n_ctx=1024 * 128,
+    generation_kwargs={"max_new_tokens": 1024 * 128},
 )
 task = ArgillaLabeller(llm=llm)
 task.load()
@@ -126,6 +127,7 @@ result = client.predict(
 interface = gr.Interface(
     fn=process_records_gradio,
     inputs=[
+        gr.Number(label="Number of Records"),
         gr.Code(label="Records (JSON)", language="json", lines=5),
         gr.Code(label="Example Records (JSON, optional)", language="json", lines=5),
         gr.Code(label="Fields (JSON, optional)", language="json"),