AI-Powered Detection ยท Hindi ยท Hinglish ยท English
Stop getting played
by the algorithm.
Paste any Indian YouTube URL. Catch-Bait analyses the title, thumbnail,
and engagement signals using a multimodal model trained on 2,066 videos from 50+ Indian channels.
""", unsafe_allow_html=True)
c1, c2 = st.columns([5, 1])
with c1:
url = st.text_input("u", label_visibility="collapsed",
placeholder="https://www.youtube.com/watch?v=...", key="u")
with c2:
go = st.button("Analyse โ")
st.markdown("""
Try:
CarryMinati
ABP News
BB Ki Vines
Dhruv Rathee
Physics Wallah
Elvish Yadav
""", unsafe_allow_html=True)
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# ANALYSIS
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
if go and url:
if not YT_KEY:
st.error("โ ๏ธ YOUTUBE_API_KEY not found in Space secrets.")
st.stop()
vid_id = extract_id(url)
if not vid_id:
st.error("Could not parse a YouTube video ID from that URL.")
st.stop()
with st.spinner("Fetching video metadataโฆ"):
meta = fetch_meta(vid_id)
if not meta:
st.error("Video not found. Check the URL or try another video.")
st.stop()
with st.spinner("Loading thumbnailโฆ"):
t_url = meta["thumb"] or f"https://img.youtube.com/vi/{vid_id}/hqdefault.jpg"
pil = Image.open(BytesIO(requests.get(t_url, timeout=12).content)).convert("RGB")
with st.spinner("Running multimodal inferenceโฆ"):
hs = h_score(meta["title"])
prob = run_predict(
meta["title"], pil, meta["views"], meta["likes"],
meta["comments"], meta["duration"], meta["tags"], hs)
is_cb = prob >= 0.5
p_cb = prob
p_safe = 1 - p_cb
st.markdown('', unsafe_allow_html=True)
# โโ VERDICT BANNER โโ
if is_cb:
st.markdown(f"""
๐ฏ
Clickbait
This title is designed to manipulate clicks
""", unsafe_allow_html=True)
else:
st.markdown(f"""
โ
Not Clickbait
This video title accurately describes its content
""", unsafe_allow_html=True)
# โโ TWO COLUMNS: Video info LEFT, Confidence + Signals RIGHT โโ
left, right = st.columns([1.2, 1])
with left:
hl = highlight(meta["title"])
found = [w for w in HINDI_BAIT if w.lower() in meta["title"].lower()]
bnote = '
๐ก highlighted = detected bait terms
' if found else ""
st.markdown(f"""
Video
{hl}
{bnote}
๐บ {meta['channel']} ยท {meta['date']}
๐ {fmt_n(meta['views'])}
๐ {fmt_n(meta['likes'])}
๐ฌ {fmt_n(meta['comments'])}
โฑ {fmt_d(meta['duration'])}
""", unsafe_allow_html=True)
with right:
# Confidence bars only
st.markdown(f"""
Prediction confidence
Not Clickbait{p_safe:.1%}
""", unsafe_allow_html=True)
# โโ SIGNAL GRID โ full width below columns โโ
caps_r = sum(1 for c in meta["title"] if c.isupper()) / max(len(meta["title"]), 1)
emoji_n = len(re.findall(r'[\U00010000-\U0010ffff]', meta["title"]))
like_r = meta["likes"] / max(meta["views"], 1)
tag_n = len([x for x in meta["tags"].split(",") if x.strip()])
def sc_html(ico, nm, val, sb, raw, hi, md):
tc = sig_top(raw, hi, md)
return f'
'
st.markdown(f"""
Signal breakdown
{sc_html("๐ค","Bait words",str(len(found)),", ".join(found[:2]) or "none",len(found),3,1)}
{sc_html("๐ ","Caps ratio",f"{caps_r:.0%}","of English chars",caps_r,0.4,0.2)}
{sc_html("๐ฎ","Emojis",str(emoji_n),"in title",emoji_n,3,1)}
{sc_html("โ","Question","Yes" if "?" in meta["title"] else "No","curiosity gap",1 if "?" in meta["title"] else 0,1,0.5)}
{sc_html("โ","Exclamation","Yes" if "!" in meta["title"] else "No","urgency signal",1 if "!" in meta["title"] else 0,1,0.5)}
{sc_html("๐","Like rate",f"{like_r:.2%}","likes / view",like_r,0.08,0.03)}
{sc_html("๐ท๏ธ","Tag count",str(tag_n),"metadata tags",0,0,-1)}
{sc_html("๐ฏ","H-score",f"{hs:.1f}/10","heuristic pre-filter",hs,6,3)}
""", unsafe_allow_html=True)
# โโ HOW IT WORKS โโ
st.markdown("""
How it works
Step 1
๐ก
Fetch
YouTube Data API v3 pulls title, thumbnail, stats, and tags live
Step 2
๐ง
Embed
MuRIL encodes text (768-d) ยท CLIP ViT-B/32 encodes thumbnail (512-d)
Step 3
๐
Fuse
L2-normalised embeddings + 12 metadata features โ 1292-d vector
Step 4
โก
Classify
XGBoost predicts probability ยท AUC-ROC 0.9952 ยท Macro F1 0.9609
""", unsafe_allow_html=True)
st.markdown('