add app file

.env

@@ -0,0 +1,2 @@
+HF_TOKEN="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"
+

app.py

@@ -0,0 +1,49 @@
+import os
+from pathlib import Path
+
+import pandas as pd
+import streamlit as st
+from datasets import load_dataset
+from dotenv import load_dotenv
+
+if Path(".env").is_file():
+    load_dotenv(".env")
+
+st.set_page_config(layout="wide")
+
+HF_TOKEN = os.getenv("HF_TOKEN")
+
+ds = load_dataset("HuggingFaceH4/instruction-model-outputs-filtered", split="train", use_auth_token=HF_TOKEN)
+
+st.markdown("# Instruction Model Outputs")
+st.markdown(
+    """This app shows the outputs of various open-souce, instruction-trained models from a [dataset](https://huggingface.co/datasets/HuggingFaceH4/instruction-model-outputs-filtered) of human demonstrations filtered for overlap with the original prompt and canned responses. Hit the button below to view a few random samples from the generated outputs."""
+)
+st.markdown(
+    """**Notes**
+* Some outputs contain a `Human:` prefix - this is likely due to the fact each model was prompted to be a dialogue agent.
+* The outputs were generated deterministically with `temperature=0` and `max_new_tokens=100`
+"""
+)
+
+button = st.button("Show me what you got!")
+
+if button is True:
+    sample_ds = ds.shuffle().select(range(5))
+
+    for sample in sample_ds:
+        st.markdown(f'**Prompt:** {sample["prompt"]}')
+
+        df = pd.DataFrame.from_records(sample["outputs"])
+
+        # CSS to inject contained in a string
+        hide_table_row_index = """
+                    <style>
+                    thead tr th:first-child {display:none}
+                    tbody th {display:none}
+                    </style>
+                    """
+
+        # Inject CSS with Markdown
+        st.markdown(hide_table_row_index, unsafe_allow_html=True)
+        st.table(df)

requirements.txt

@@ -0,0 +1,2 @@
+datasets
+python-dotenv