Adding examples to iface. Returning year field to df earlier.

app.py

@@ -241,15 +241,15 @@ def predict_gender_pronouns(
 
     results = pd.concat(dfs, axis=1).set_index("year")
 
-    female_df = results.filter(regex=".*f_")
+    female_df = results.filter(regex=".*f_").reset_index()  # Gradio doesn't 'see' index?
     female_df_for_plot = (
-        female_df.reset_index()
-    )  # Gradio timeseries requires x-axis as column?
+        female_df
+    ) 
 
-    male_df = results.filter(regex=".*m_")
+    male_df = results.filter(regex=".*m_").reset_index()  # Gradio doesn't 'see' index?
     male_df_for_plot = (
-        male_df.reset_index()
-    )  # Gradio timeseries requires x-axis as column?
+        male_df
+    )  
 
     return (
         target_text_w_masks,
@@ -263,24 +263,17 @@ def predict_gender_pronouns(
 title = "Changing Gender Pronouns"
 description = """
 <h2> Intro </h2>
-
 This is a demo for a project exploring possible spurious correlations that have been learned by our models. We first examined the training datasets and learning tasks to hypothesize what spurious correlations may exist. Below we can condition on these variables to determine what effect they may have on the prediction outcomes.
-
 Specially in this demo: In a user provided sentence, with at least one reference to a `DATE` and one gender pronoun, we will see how sweeping through a range of `DATE` values can change the predicted pronouns. This effect can be observed in BERT base models and in our fine-tuned models (with a specific pronoun predicting task on the [wiki-bio](https://huggingface.co/datasets/wiki_bio) dataset).
-
 One way to explain this phenomenon is by looking at a likely data generating process for biographical-like data in both the main BERT training dataset as well as the `wiki_bio` dataset, in the form of a causal DAG.
  
 <h2> Causal DAG </h2> 
-
 In the DAG, we can see that `birth_place`, `birth_date` and `gender` are all independent elements that have no common cause with the other covariates in the DAG. However `birth_place`, `birth_date` and `gender` may all have a role in causing one's `access_to_resources`, with the general trend that `access_to_resources` has become less gender-dependent over time, but not in every `birth_place`, with recent events in Afghanistan providing a stark counterexample to this trend. 
-
 Importantly, `access_to_resources` determines how, **if at all**, you may appear in the dataset's `context_words`.
-
 We argue that although there are complex causal interactions between each word in any given sentence, the `context_words` are more likely to cause the `gender_pronouns`, rather than vice versa. For example, if the subject is a famous doctor and the object is her wealthy father, these context words will determine which person is being referred to, and thus which gendered-pronoun to use.
  
  
 In this graph, arrow heads are intended to show the assumed direction of causation. E.g. as described above, we are claiming `context_words` cause the `gender_pronouns`. While causation follow direction of the arrows, statistical correlation can flow in any direction (it is cause-agnostic). 
-
 In the case of this graph, any pink path between `context_words` and `gender_pronouns` will allow the flow of statistical correlation, inviting confounding and thus spurious correlations into the trained model.
  
 <center>
@@ -291,7 +284,6 @@ In the case of this graph, any pink path between `context_words` and `gender_pro
 Those familiar with causal DAGs may note when can simply condition on `gender` to block any confounding between the `context_words` and the `gender_pronouns`.  However, this is not always possible, particularly in generative or mask-filling tasks where gender may be unknown, common in language modeling and in the demo below.  
  
  <h2> How to use this demo </h2> 
-
 In this demo, a user can add any sentence that contains at least one gender pronoun and the capitalized word `DATE`. We then sweep through a range of `date` values in the place of `DATE`, while masking (for prediction) the gender pronouns (included in the list below).
 ```
 gendered_lists = [
@@ -304,26 +296,27 @@ gendered_lists = [
    ["husband", "wife"],
 ]
 ```
-
 In addition to choosing the test sentence, we ask that you pick how the fine-tuned model was trained:
 - conditioning variable: which, if any, conditioning variable from the three noted above in the DAG, was included in the text at train time.
 - loss function weight: weight assigned to the minority class (female pronouns in this fine-tuning dataset) that was included in the text at train time.
-
 You can also optionally pick a bert-like model for comparison.
 
- <h2> What are the results</h2> 
 
+Some notes: 
+- Gradio currently only supports 6 plotting colors (but there are [plans](https://github.com/gradio-app/gradio/issues/1088) to support more!), so best to not select too many models at once for now.
+- If the dataframes appear to not update with new fields, it may help to 'Clear' the fields before 'Submitting' new inputs. 
+
+
+ <h2> What are the results</h2> 
 In the resulting plots, we can look for a dose-response relationship between:
 - our treatment: the sample text,
 - and our outcome: the predicted gender of pronouns in the text.
  
 Specifically, we are seeing if 1) making larger magnitude intervention: an older `DATE` in the text will, 2) result in a larger magnitude effect in the outcome: higher percentage of predicted female pronouns.
-
 Some trends that appear in the test sentences I have tried:
 - Conditioning on `birth_date` metadata in both training and inference text has the largest dose-response relationship. This seems reasonable, as the fine-tuned model is able to 'stratify' a learned relationship between gender pronouns and dates, when both are present in the text.
 - While conditioning on either no metadata or `birth_place` data training, have similar middle-ground effects for this inference task. 
 - Finally, conditioning on `name` metadata in training, (while again conditioning on `date` in inference) has almost no dose-response relationship. It appears the learning of a `name —> gender pronouns` relationship was sufficiently successful to overwhelm any potential more nuanced learning, such as that driven by `birth_date` or `place`. 
-
 Please feel free to ping me on the Hugging Face discord (I'm 'emily_learner' there), with any feedback/comments/concerns or interesting findings!
 """
 
@@ -331,6 +324,30 @@ Please feel free to ping me on the Hugging Face discord (I'm 'emily_learner' the
 article = "Check out [main colab notebook](https://colab.research.google.com/drive/14ce4KD6PrCIL60Eng-t79tEI1UP-DHGz?usp=sharing#scrollTo=Mg1tUeHLRLaG) \
  with a lot more details about this method and implementation."
 
+ceo_example = [
+    20,
+    ["none", "birth_date", "name"],
+    FEMALE_WEIGHTS,
+    [],
+    'Born in DATE, she was a CEO. Her work was greatly respected, and she was well-regarded in her field.',
+]
+
+death_date_example = [
+    10,
+    ['birth_date'],
+    [1.5],
+    BERT_LIKE_MODELS,
+    'Died in DATE, she was reconigized for her great accomplishments to the field of teaching.'
+]
+
+building_date_example = [
+    30,
+    ['birth_date'],
+    [1.5],
+    BERT_LIKE_MODELS,
+    'Built in DATE, her building provided the perfect environment for her job as a teacher.'
+]
+
 gr.Interface(
     fn=predict_gender_pronouns,
     inputs=[
@@ -385,4 +402,5 @@ gr.Interface(
     title=title,
     description=description,
     article=article,
+    examples=[ceo_example, death_date_example, building_date_example]
 ).launch()