Update app.py

app.py

@@ -324,15 +324,15 @@ def complete(sample, k, current_example):
     output = prefix
     for token in client.text_generation(prefix, do_sample=False, max_new_tokens=512, stream=True):
         if token == "<|endoftext|>":
-            bleu_score = {"Memorization score (BLEU)": bleu.compute(predictions=[output],
+            bleu_score = {"BLEU": bleu.compute(predictions=[output],
                                 references=[current_example])["bleu"]}
             return diff(output, current_example), gr.Label.update(value=bleu_score), current_example
         output += token
-        bleu_score = {"Memorization score (BLEU)": bleu.compute(predictions=[output],
+        bleu_score = {"BLEU": bleu.compute(predictions=[output],
                                 references=[current_example])["bleu"]}
         yield diff(output, current_example), gr.Label.update(value=bleu_score), current_example
         # yield output, diff_texts(output, sample), gr.Label.update(value=bleu_score)
-    bleu_score = {"Memorization score (BLEU)": bleu.compute(predictions=[output],
+    bleu_score = {"BLEU": bleu.compute(predictions=[output],
                                 references=[current_example])["bleu"]}
     # return output, diff_texts(output, sample), gr.Label.update(value=bleu_score)
     return diff(output, current_example), gr.Label.update(value=bleu_score), current_example
@@ -363,7 +363,7 @@ DEFAULT_SAMPLE_PREFIX = tokenizer.decode(tokenizer(DEFAULT_SAMPLE)["input_ids"][
 
 style = theme.Style()
 
-with gr.Blocks(theme=style) as demo:
+with gr.Blocks(theme=style, css=modifs) as demo:
     current_example = gr.State(value=DEFAULT_SAMPLE)
     with gr.Column():
         gr.Markdown(title)
@@ -378,7 +378,8 @@ with gr.Blocks(theme=style) as demo:
                     id="instruction",
                     placeholder="Output",
                     lines=5,
-                    label="Prompt",
+                    label="Training sample",
+                    info="This is an example from The Stack dataset."
                     value=DEFAULT_SAMPLE_PREFIX,
                     disable=True,
                     interactive=False,
@@ -392,19 +393,18 @@ with gr.Blocks(theme=style) as demo:
                             You can reduce the Prefix size in the Advanced parameters to reduce the context length and see if the model still extracts the training sample.""")    
         with gr.Row():
             with gr.Column():
-                with gr.Accordion("Prompt size", open=True):
+                with gr.Accordion("Prefix size", open=True):
                     k = gr.Slider(minimum=1, maximum=DEFAULT_SAMPLE_MAX_TOKENS, value=DEFAULT_K,
                                   step=1,
-                                  label="Prompt size",
-                                  info="""Number of tokens used in the prompt. 
-                                  Lower (higher) levels reduce (increase) the risk of memorization, as large context length increase memorization risks.""")
-                submit = gr.Button("Check", variant="primary")
+                                  info="""Number of tokens used to assess memorization. 
+                                  The more tokens are used, the more likely one can observe the LLM finishing the prompt with the verbatim code used in the training set.""")
+                submit = gr.Button("Check memorization", variant="primary")
                 examples_dropdown = gr.Dropdown(choices=list(examples.keys()), value=list(examples.keys())[0],
                                     interactive=True,
-                                    label="Training set samples")
+                                    label="Training set samples",
+                                    info="""You can choose among high/low memorization examples from The Stack.
+                                    More samples are available below.""")
             with gr.Column():
-                # with gr.Row():
-                #    output = gr.Textbox(lines=5, label="Completion", interactive=False)
                 diff_HTML = gr.HTML(
                     label="Diff")