updates

app.py

@@ -4,14 +4,28 @@ import corpus
 from PIL import Image
 
 PAGES = {
-    "Single Explanation": single,
-    "Corpus Explanation": corpus,
+    "Instance Explanation": single,
+    "Dataset Explanation": corpus,
 }
 
-# st.sidebar.title("Explore ferret!")
-
 logo = Image.open("static/logo.png")
 st.sidebar.image(logo)
+st.sidebar.title("*ferret* showcase")
 
+st.sidebar.markdown(
+    """
+    Welcome to the *ferret* showcase!
+    
+    You will find two main functionalities.
+    """
+)
 page = st.sidebar.radio("", list(PAGES.keys()))
+
+st.sidebar.markdown(
+    """
+    In the single-*instance* page, you can evaluate our built-in explainers on your favourite model.
+    Choosing *dataset* mode, you will evaluate explainers on state-of-the-art datasets from our Dataset API.
+    """
+)
+
 PAGES[page].body()

requirements.txt

@@ -1,3 +1,3 @@
 transformers==4.20.1
-ferret-xai==0.2.4
+ferret-xai==0.3.0
 sentencepiece

single.py

@@ -23,14 +23,12 @@ def get_tokenizer(tokenizer_name):
 
 def body():
 
+    st.title("Explain individual texts")
+
     st.markdown(
         """
-        # Welcome to the *ferret* showcase
-
         You are working now on the *single instance* mode -- i.e., you will work and
         inspect one textual query at a time.
-
-        ## Sentiment Analysis
  
         Post-hoc explanation techniques discose the rationale behind a given prediction a model
         makes while detecting a sentiment out of a text. In a sense the let you *poke* inside the model.
@@ -52,12 +50,15 @@ def body():
     col1, col2 = st.columns([3, 1])
     with col1:
         model_name = st.text_input("HF Model", DEFAULT_MODEL)
+        config = AutoConfig.from_pretrained(model_name)
+
     with col2:
+        class_labels = list(config.id2label.values())
         target = st.selectbox(
             "Target",
-            options=range(5),
+            options=class_labels,
             index=1,
-            help="Positional index of your target class.",
+            help="Class label you want to explain.",
         )
 
     text = st.text_input("Text", "I love your style!")
@@ -69,7 +70,6 @@ def body():
         with st.spinner("Preparing the magic. Hang in there..."):
             model = get_model(model_name)
             tokenizer = get_tokenizer(model_name)
-            config = get_config(model_name)
             bench = Benchmark(model, tokenizer)
 
         st.markdown("### Prediction")
@@ -80,26 +80,29 @@ def body():
         st.text(scores_str)
 
         with st.spinner("Computing Explanations.."):
-            explanations = bench.explain(text, target=target)
+            explanations = bench.explain(text, target=class_labels.index(target))
 
         st.markdown("### Explanations")
         st.dataframe(bench.show_table(explanations))
+        st.caption("Darker red (blue) means higher (lower) contribution.")
 
         with st.spinner("Evaluating Explanations..."):
             evaluations = bench.evaluate_explanations(
-                explanations, target=target, apply_style=False
+                explanations, target=class_labels.index(target), apply_style=False
             )
 
         st.markdown("### Faithfulness Metrics")
         st.dataframe(bench.show_evaluation_table(evaluations))
+        st.caption("Darker colors mean better performance.")
 
         st.markdown(
             """
             **Legend**
 
             - **AOPC Comprehensiveness** (aopc_compr) measures *comprehensiveness*, i.e., if the explanation captures all the tokens needed to make the prediction. Higher is better.
-            - **AOPC Sufficiency** (aopc_suff) measures *sufficiency*, i.e., if the relevant tokens in the explanation are sufficient to make the prediction. Lower is better.
             
+            - **AOPC Sufficiency** (aopc_suff) measures *sufficiency*, i.e., if the relevant tokens in the explanation are sufficient to make the prediction. Lower is better.
+
             - **Leave-On-Out TAU Correlation** (taucorr_loo) measures the Kendall rank correlation coefficient τ between the explanation and leave-one-out importances. Closer to 1 is better. 
             
             See the paper for details.
@@ -111,3 +114,23 @@ def body():
         # It is computed as the drop in the model probability if only the relevant tokens of the explanations are considered. The lower the sufficiency, the more faithful is the explanation since there is less change in the model prediction.
 
         # The latter are computed by omittig individual input tokens and measuring the variation on the model prediction. The closer the τ correlation is to 1, the more faithful is the explanation.
+
+        st.markdown(
+            """
+            **In code, it would be as simple as**
+        """
+        )
+        st.code(
+            f"""
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+from ferret import Benchmark
+
+model = AutoModelForSequenceClassification.from_pretrained("{model_name}")
+tokenizer = AutoTokenizer.from_pretrained("{model_name}")
+
+bench = Benchmark(model, tokenizer)
+explanations = bench.explain("{text}")
+evaluations = bench.evaluate_explanations(explanations)
+            """,
+            language="python",
+        )