low-margin uncertainty detection
Browse files- src/streamlit_app.py +28 -7
src/streamlit_app.py
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
import os
|
| 2 |
|
| 3 |
-
# ✅
|
| 4 |
os.environ["TRANSFORMERS_CACHE"] = "/tmp/hf"
|
| 5 |
os.environ["HF_HOME"] = "/tmp/hf"
|
| 6 |
os.environ["STREAMLIT_HOME"] = "/tmp/.streamlit"
|
|
@@ -10,14 +10,15 @@ os.makedirs("/tmp/.streamlit", exist_ok=True)
|
|
| 10 |
import streamlit as st
|
| 11 |
from transformers import pipeline
|
| 12 |
|
| 13 |
-
# 🔐 Load
|
| 14 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
| 15 |
|
| 16 |
# 🔁 Load the private model using the token
|
| 17 |
classifier = pipeline(
|
| 18 |
"text-classification",
|
| 19 |
model="azratuni/isl-classifier",
|
| 20 |
-
token=HF_TOKEN
|
|
|
|
| 21 |
)
|
| 22 |
|
| 23 |
# 🧠 Streamlit UI
|
|
@@ -30,11 +31,31 @@ text_input = st.text_area("Input text", placeholder="e.g. Muslims are terrorists
|
|
| 30 |
if st.button("Classify"):
|
| 31 |
if text_input.strip():
|
| 32 |
with st.spinner("Classifying..."):
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
|
| 37 |
st.success(f"**Prediction:** {label}")
|
| 38 |
-
st.write(f"**Confidence:** {
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 39 |
else:
|
| 40 |
st.warning("Please enter some text.")
|
|
|
|
| 1 |
import os
|
| 2 |
|
| 3 |
+
# ✅ Fix: Use writable directories for cache and metrics
|
| 4 |
os.environ["TRANSFORMERS_CACHE"] = "/tmp/hf"
|
| 5 |
os.environ["HF_HOME"] = "/tmp/hf"
|
| 6 |
os.environ["STREAMLIT_HOME"] = "/tmp/.streamlit"
|
|
|
|
| 10 |
import streamlit as st
|
| 11 |
from transformers import pipeline
|
| 12 |
|
| 13 |
+
# 🔐 Load Hugging Face token from Space secrets
|
| 14 |
HF_TOKEN = os.getenv("HF_TOKEN")
|
| 15 |
|
| 16 |
# 🔁 Load the private model using the token
|
| 17 |
classifier = pipeline(
|
| 18 |
"text-classification",
|
| 19 |
model="azratuni/isl-classifier",
|
| 20 |
+
token=HF_TOKEN,
|
| 21 |
+
return_all_scores=True # ✅ Return all class scores
|
| 22 |
)
|
| 23 |
|
| 24 |
# 🧠 Streamlit UI
|
|
|
|
| 31 |
if st.button("Classify"):
|
| 32 |
if text_input.strip():
|
| 33 |
with st.spinner("Classifying..."):
|
| 34 |
+
scores = classifier(text_input)[0]
|
| 35 |
+
scores = sorted(scores, key=lambda x: x["score"], reverse=True)
|
| 36 |
+
|
| 37 |
+
top_label = scores[0]["label"]
|
| 38 |
+
top_score = scores[0]["score"]
|
| 39 |
+
second_score = scores[1]["score"]
|
| 40 |
+
margin = top_score - second_score
|
| 41 |
+
|
| 42 |
+
label_map = {
|
| 43 |
+
"LABEL_0": "Not Islamophobic",
|
| 44 |
+
"LABEL_1": "Islamophobic"
|
| 45 |
+
}
|
| 46 |
+
|
| 47 |
+
label = label_map.get(top_label, top_label)
|
| 48 |
|
| 49 |
st.success(f"**Prediction:** {label}")
|
| 50 |
+
st.write(f"**Confidence:** {top_score:.2%}")
|
| 51 |
+
|
| 52 |
+
# ⚠️ Uncertainty warning if margin is low
|
| 53 |
+
if margin < 0.15:
|
| 54 |
+
st.warning("⚠️ The model is uncertain. Both classes received similar confidence scores.")
|
| 55 |
+
|
| 56 |
+
# 🧾 Optional: show both scores
|
| 57 |
+
st.markdown("**Score breakdown:**")
|
| 58 |
+
for s in scores:
|
| 59 |
+
st.write(f"- {label_map.get(s['label'], s['label'])}: {s['score']:.2%}")
|
| 60 |
else:
|
| 61 |
st.warning("Please enter some text.")
|