Upload folder using huggingface_hub

app.py

@@ -67,7 +67,7 @@ CSS = """
 HERO = """<div id="hero"><div class="in">
 <h1>اردو <span class="a">Education & Reasoning</span></h1>
 <p>A Gemma-3-4B model adapted to Urdu by translating English knowledge corpora into Urdu with
-<b>Adaption AutoScientist</b> — and benchmarked on UrduMMLU.</p>
+<b>Adaption AutoScientist</b>, then benchmarked on UrduMMLU.</p>
 <div class="badges"><span class="b">Gemma-3-4B</span><span class="b">Adaption AutoScientist</span>
 <span class="b">UrduMMLU 46.2%</span><span class="b">+1.3 vs base</span></div></div></div>"""
 
@@ -76,20 +76,26 @@ KPI = """<div class="kpi">
 <div class="c"><div class="v">+3.6</div><div class="l">Profession</div></div>
 <div class="c"><div class="v">+2.7</div><div class="l">Social Sci.</div></div>
 <div class="c"><div class="v">46.2%</div><div class="l">UrduMMLU overall</div></div>
-<div class="c"><div class="v">beats</div><div class="l">base Gemma (44.96%)</div></div></div>"""
+<div class="c"><div class="v">+1.3</div><div class="l">vs base overall</div></div></div>"""
 
-APPROACH = """<div class="sec"><h2>How it was built</h2>
-<p>Most UrduMMLU domains test <b>transferable knowledge</b> (science, math, social science, reasoning) that
-exists in abundant English datasets. We assembled ~40K examples from open English corpora
-(MMLU, GSM8K, MATH, ARC, AQuA, commonsense) plus native-Urdu instruction data, then used
-<b>Adaption AutoScientist</b> and the <b>Adaptive Data</b> pipeline to adapt each example into natural
-Pakistani Urdu — adding richer prompts and English reasoning traces — and supervised-fine-tuned Gemma-3-4B on the result.</p></div>"""
+VALIDATED = """<div class="sec"><h2>What this validates</h2>
+<p>We tested whether adapting English knowledge corpora into Urdu with Adaption AutoScientist improves a 4B
+model on a native Urdu benchmark. It does, for knowledge that is language-independent: every such domain
+improved and the model exceeded its base overall (44.96% to 46.21%). The effect does not extend to Urdu
+literature, which is intrinsic to the language and requires native data rather than translation.</p></div>"""
 
-FINDING = """<div class="note">📌 <b>The honest finding.</b> Cross-lingual adaptation lifted <i>every transferable
-domain</i> and beat the base model. The one domain that didn't improve was <b>Urdu literature / Humanities</b>
-(−2.5) — knowledge that is intrinsic to the Urdu language and <i>cannot</i> be sourced by adapting English.
-Closing it needs <b>better native Urdu-literature datasets</b>, not better adaptation — a data-availability
-problem, not a method problem.</div>"""
+APPROACH = """<div class="sec"><h2>Method</h2>
+<p>Most UrduMMLU subjects test knowledge that is largely language-independent: science, mathematics,
+reasoning, and social studies. We assembled about 40,000 examples from open English datasets covering these
+subjects, together with native-Urdu instruction and literature data, then used <b>Adaption AutoScientist</b>
+and the <b>Adaptive Data</b> pipeline to translate and localise each example into Pakistani Urdu, adding a
+reformulated prompt and an English reasoning trace. Gemma-3-4B was supervised-fine-tuned on the result and
+evaluated on UrduMMLU, zero-shot.</p></div>"""
+
+FINDING = """<div class="note"><b>Boundary of the method.</b> Cross-lingual adaptation improved the science,
+mathematics, reasoning, and social-knowledge domains, but Urdu literature declined by 2.5 points. That
+content cannot be produced by translating English sources; improving it requires native Urdu literary data.
+This is a limitation of available data, not of the adaptation method.</div>"""
 
 FOOT = """<div id="foot">Model: <a href="https://huggingface.co/abdullah693/gemma-3-4b-it-urdu-edu-reasoning" target="_blank">abdullah693/gemma-3-4b-it-urdu-edu-reasoning</a>
  · Adapted with <b>Adaption AutoScientist</b> · Eval: <a href="https://huggingface.co/datasets/MBZUAI/UrduMMLU" target="_blank">UrduMMLU</a> (Urdu, 0-shot)<br>
@@ -111,13 +117,19 @@ with gr.Blocks(title="Urdu Education & Reasoning", theme=gr.themes.Soft(primary_
         mt = gr.Slider(64, 512, value=256, step=32, label="Max new tokens")
         tp = gr.Slider(0.0, 1.0, value=0.0, step=0.1, label="Temperature")
     gr.ChatInterface(respond, additional_inputs=[mt, tp], examples=[[e] for e in EXAMPLES], cache_examples=False)
-    # ── results below the chat ──
-    gr.HTML('<div class="sec" style="margin-top:18px"><h2>The result</h2></div>')
+    # ── what we validated ──
+    gr.HTML('<div style="margin-top:18px"></div>')
+    gr.HTML(VALIDATED)
     gr.HTML(KPI)
+    gr.Image("assets/fig_validated.png", show_label=False, container=False)
+    # ── results ──
+    gr.HTML('<div class="sec" style="margin-top:6px"><h2>Results on UrduMMLU</h2></div>')
     with gr.Row():
         gr.Image("assets/fig_overall.png", show_label=False, container=False)
         gr.Image("assets/fig_domains.png", show_label=False, container=False)
+    # ── method ──
     gr.HTML(APPROACH)
+    gr.Image("assets/fig_pipeline.png", show_label=False, container=False)
     gr.Image("assets/fig_data.png", show_label=False, container=False)
     gr.HTML(FINDING)
     gr.HTML(FOOT)