Update app.py

app.py

@@ -18,11 +18,18 @@ Options:
 - greedy: accurate completions (may be more boring or have repetitions)
 """
 
-wip_description = """JAX / Flax Gradio Demo for BLOOM. The 176B BLOOM model running on a TPU v3-256 pod, with 2D model parallelism and custom mesh axes.
+wip_description = """Gradio Demo for JAX / Flax BLOOM. The 176B BLOOM model running on a TPU v3-256 pod, with 2D model parallelism and custom mesh axes.
 
-Note:
-1. For this WIP demo, only **sampling** is supported.
-2. Rendering of the screenshot is currently not optimised. To experience the true speed of JAX / Flax, tick 'just output raw text'.
+
+Note: rendering of the screenshot is currently not optimised. To experience the true speed of JAX / Flax, tick 'just output raw text'.
+
+Tips:
+- Do NOT talk to BLOOM as an entity, it's not a chatbot but a webpage/blog/article completion model.
+- For the best results: MIMIC a few sentences of a webpage similar to the content you want to generate.
+Start a paragraph as if YOU were writing a blog, webpage, math post, coding article and BLOOM will generate a coherent follow-up. Longer prompts usually give more interesting results.
+Options:
+- sampling: imaginative completions (may be not super accurate e.g. math/history)
+- greedy: accurate completions (may be more boring or have repetitions)
 """
 
 API_URL = os.getenv("API_URL")
@@ -33,7 +40,7 @@ examples = [
         64, "sampling", True],
     ['A poem about the beauty of science by Alfred Edgar Brittle\nTitle: The Magic Craft\nIn the old times', 64,
      "sampling", True],
-    ['استخراج العدد العاملي في لغة بايثون:', 64, "greedy", True],
+    ['استخراج العدد العاملي في لغة بايثون:', 64, "sampling", True],
     ["Pour déguster un ortolan, il faut tout d'abord", 64, "sampling", True],
     [
         'Traduce español de España a español de Argentina\nEl coche es rojo - el auto es rojo\nEl ordenador es nuevo - la computadora es nueva\nel boligrafo es negro -',
@@ -41,7 +48,7 @@ examples = [
     [
         'Estos ejemplos quitan vocales de las palabras\nEjemplos:\nhola - hl\nmanzana - mnzn\npapas - pps\nalacran - lcrn\npapa -',
         64, "sampling", True],
-    ["Question: If I put cheese into the fridge, will it melt?\nAnswer:", 64, "greedy", True],
+    ["Question: If I put cheese into the fridge, will it melt?\nAnswer:", 64, "sampling", True],
     ["Math exercise - answers:\n34+10=44\n54+20=", 64, "sampling", True],
     [
         "Question: Where does the Greek Goddess Persephone spend half of the year when she is not with her mother?\nAnswer:",
@@ -60,10 +67,9 @@ def query(payload):
 
 
 def inference(input_sentence, max_length, sample_or_greedy, raw_text=True):
-    do_sample = sample_or_greedy == "sampling"
     payload = {
         "inputs": input_sentence,
-        "do_sample": do_sample,
+        "do_sample": True,
         # "max_new_tokens": max_length
     }
 
@@ -106,7 +112,7 @@ def inference(input_sentence, max_length, sample_or_greedy, raw_text=True):
         init_font_size=142,
         right_align=False,
     )
-    return img, data[0]['generated_text'][0][0]
+    return img, data[0]['generated_text'][0]
 
 
 gr.Interface(
@@ -114,7 +120,7 @@ gr.Interface(
     [
         gr.inputs.Textbox(label="Input"),
         gr.inputs.Radio([64], default=64, label="Tokens to generate"),
-        gr.inputs.Radio(["sampling", "greedy"], label="Sample or greedy", default="sampling"),
+        gr.inputs.Radio(["sampling"], label="Sample or greedy", default="sampling"),
         gr.Checkbox(label="Just output raw text", value=True),
     ],
     ["image", "text"],