bold settings name descriptions

demo_watermark.py

@@ -579,26 +579,26 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
             """
             #### Generation Parameters:
 
-            - Decoding Method : We can generate tokens from the model using either multinomial sampling or we can generate using greedy decoding.
-            - Sampling Temperature : If using multinomial sampling we can set the temperature of the sampling distribution. 
+            - **Decoding Method** : We can generate tokens from the model using either multinomial sampling or we can generate using greedy decoding.
+            - **Sampling Temperature** : If using multinomial sampling we can set the temperature of the sampling distribution. 
                                 0.0 is equivalent to greedy decoding, and 1.0 is the maximum amount of variability/entropy in the next token distribution.
                                 0.7 strikes a nice balance between faithfulness to the model's estimate of top candidates while adding variety. Does not apply for greedy decoding.
-            - Generation Seed : The integer to pass to the torch random number generator before running generation. Makes the multinomial sampling strategy
+            - **Generation Seed** : The integer to pass to the torch random number generator before running generation. Makes the multinomial sampling strategy
                                 outputs reproducible. Does not apply for greedy decoding.
-            - Number of Beams : When using greedy decoding, we can also set the number of beams to > 1 to enable beam search. 
+            - **Number of Beams** : When using greedy decoding, we can also set the number of beams to > 1 to enable beam search. 
                                 This is not implemented/excluded from paper for multinomial sampling but may be added in future.
-            - Max Generated Tokens : The `max_new_tokens` parameter passed to the generation method to stop the output at a certain number of new tokens. 
+            - **Max Generated Tokens** : The `max_new_tokens` parameter passed to the generation method to stop the output at a certain number of new tokens. 
                                     Note that the model is free to generate fewer tokens depending on the prompt. 
                                     Implicitly this sets the maximum number of prompt tokens possible as the model's maximum input length minus `max_new_tokens`,
                                     and inputs will be truncated accordingly.
             
             #### Watermark Parameters:
 
-            - gamma : The fraction of the vocabulary to be partitioned into the greenlist at each generation step. 
+            - **gamma** : The fraction of the vocabulary to be partitioned into the greenlist at each generation step. 
                      Smaller gamma values create a stronger watermark by enabling the watermarked model to achieve 
                      a greater differentiation from human/unwatermarked text because it is preferentially sampling 
                      from a smaller green set making those tokens less likely to occur by chance.
-            - delta : The amount of positive bias to add to the logits of every token in the greenlist 
+            - **delta** : The amount of positive bias to add to the logits of every token in the greenlist 
                         at each generation step before sampling/choosing the next token. Higher delta values 
                         mean that the greenlist tokens are more heavily preferred by the watermarked model
                         and as the bias becomes very large the watermark transitions from "soft" to "hard". 
@@ -607,7 +607,7 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
 
             #### Detector Parameters:
             
-            - z-score threshold : the z-score cuttoff for the hypothesis test. Higher thresholds (such as 4.0) make
+            - **z-score threshold** : the z-score cuttoff for the hypothesis test. Higher thresholds (such as 4.0) make
                                 _false positives_ (predicting that human/unwatermarked text is watermarked) very unlikely
                                 as a genuine human text with a significant number of tokens will almost never achieve 
                                 that high of a z-score. Lower thresholds will capture more _true positives_ as some watermarked
@@ -615,11 +615,11 @@ def run_gradio(args, model=None, device=None, tokenizer=None):
                                 be flagged as "watermarked". However, a lowere threshold will increase the chance that human text 
                                 that contains a slightly higher than average number of green tokens is erroneously flagged. 
                                 4.0-5.0 offers extremely low false positive rates while still accurately catching most watermarked text.
-            - Ignore Bigram Repeats : This alternate detection algorithm only considers the unique bigrams in the text during detection, 
+            - **Ignore Bigram Repeats** : This alternate detection algorithm only considers the unique bigrams in the text during detection, 
                                     computing the greenlists based on the first in each pair and checking whether the second falls within the list.
                                     This means that `T` is now the unique number of bigrams in the text, which becomes less than the total
                                     number of tokens generated if the text contains a lot of repetition. See the paper for a more detailed discussion.
-            - Normalizations : we implement a few basic normaliations to defend against various adversarial perturbations of the
+            - **Normalizations** : we implement a few basic normaliations to defend against various adversarial perturbations of the
                                 text analyzed during detection. Currently we support converting all chracters to unicode, 
                                 replacing homoglyphs with a canonical form, and standardizing the capitalization. 
                                 See the paper for a detailed discussion of input normalization.