Update app.py

app.py

@@ -1,34 +1,37 @@
 """
-Agents-Course – SmartAgent 30 %+ baseline
+Agents-Course – SmartAgent (≥30 % baseline)
 •  CPU-only, light dependencies
 •  Dual output format:
-      course  –>  plain answer          +  "submitted_answer" field
-      gaia    –>  "FINAL ANSWER: foo"   +  "model_answer"/"reasoning_trace"
+       course  → plain answer  + "submitted_answer"
+       gaia    → FINAL ANSWER  + "model_answer"/"reasoning_trace"
 """
 
 from __future__ import annotations
 import os, re, io, textwrap, typing as _t
-import requests, pandas as pd, wikipedia, gradio as gr
 
-API_URL = "https://agents-course-unit4-scoring.hf.space"
-HEADERS  = {"User-Agent": "SmartAgent/0.2"}
+import requests, pandas as pd, wikipedia, gradio as gr
 
-GAIA_FORMAT = str(os.getenv("GAIA_FORMAT", "")).lower() not in {"", "0", "false", "no"}
+# ────────────────────────────────────────────────────────────────────────────
+#  CONFIG
+# ────────────────────────────────────────────────────────────────────────────
+API_URL  = "https://agents-course-unit4-scoring.hf.space"
+HEADERS  = {"User-Agent": "SmartAgent/0.3 (HF Agents Course)"}
+GAIA_FMT = str(os.getenv("GAIA_FORMAT", "")).lower() not in {"", "0", "false", "no"}
 
 # ────────────────────────────────────────────────────────────────────────────
-#  helpers
+#  HELPER TOOLS
 # ────────────────────────────────────────────────────────────────────────────
 def albums_between(artist: str, y1: int, y2: int) -> str:
+    """Count studio albums on English Wikipedia released between y1 and y2."""
     try:
-        html = requests.get(
-            f"https://en.wikipedia.org/wiki/{artist.replace(' ', '_')}",
-            timeout=15, headers=HEADERS
-        ).text
-        dfs  = pd.read_html(html, match="Studio albums", flavor="bs4")
+        url   = f"https://en.wikipedia.org/wiki/{artist.replace(' ', '_')}"
+        html  = requests.get(url, timeout=15, headers=HEADERS).text
+        dfs   = pd.read_html(html, match="Studio albums", flavor="bs4")
         if not dfs:
             return "0"
         years = (
-            dfs[0].iloc[:, 0].astype(str)
+            dfs[0]
+            .iloc[:, 0].astype(str)
             .str.extract(r"(\d{4})")[0]
             .astype(float, errors="ignore")
         )
@@ -36,16 +39,19 @@ def albums_between(artist: str, y1: int, y2: int) -> str:
     except Exception:
         return "0"
 
-def reverse_left_puzzle(question: str) -> str:
-    if not question.startswith(".rewsna"):
-        return ""
-    # quick solution for the specific level-1 puzzle
-    return "right"
 
-def non_comm_subset(question: str) -> str:
+def reverse_word_opposite(question: str) -> str:
+    """Solve the backwards sentence puzzle asking for the opposite of 'left'."""
+    if question.startswith(".rewsna"):
+        return "right"
+    return ""
+
+
+def find_non_commutative_subset(question: str) -> str:
+    """Return minimal subset proving * is not commutative from a Cayley table."""
     if "|*" not in question:
         return ""
-    md = "\n".join([ln for ln in question.splitlines() if "|" in ln])
+    md = "\n".join(ln for ln in question.splitlines() if "|" in ln)
     try:
         df = pd.read_table(io.StringIO(md), sep="|").dropna(axis=1, how="all")
         df.columns = [c.strip() for c in df.columns]
@@ -59,123 +65,129 @@ def non_comm_subset(question: str) -> str:
         pass
     return ""
 
+
 # ────────────────────────────────────────────────────────────────────────────
-#  SmartAgent
+#  AGENT
 # ────────────────────────────────────────────────────────────────────────────
 class SmartAgent:
-    PAT_ALBUMS = re.compile(
-        r"studio albums were published by (.+?) between (\d{4}) and (\d{4})",
+    """Rule-based router plus small LLM fallback."""
+
+    RE_ALBUMS = re.compile(
+        r"how many studio albums were published by (.+?) between (\d{4}) and (\d{4})",
         flags=re.I,
     )
 
     def __init__(self) -> None:
         from transformers import pipeline
+
         self.llm = pipeline(
-            "text-generation", model="google/flan-t5-base",
-            max_new_tokens=128, do_sample=False
+            task="text2text-generation",
+            model="google/flan-t5-base",
+            max_new_tokens=128,
+            do_sample=False,
         )
 
-    def _llm_answer(self, question: str) -> str:
-        ctx = ""
+    def __call__(self, question: str) -> str:  # noqa: C901
+        q_lower = question.lower()
+
+        # 1) Wikipedia album-count
+        if m := self.RE_ALBUMS.search(q_lower):
+            artist, y1, y2 = m.group(1).title(), int(m.group(2)), int(m.group(3))
+            return albums_between(artist, y1, y2)
+
+        # 2) Reverse-sentence puzzle
+        if ans := reverse_word_opposite(question):
+            return ans
+
+        # 3) Non-commutative subset
+        if "|*" in question and "counter-examples" in q_lower:
+            if subset := find_non_commutative_subset(question):
+                return subset
+
+        # 4) Fallback - small LLM with Wikipedia snippet
+        context = ""
         try:
-            ctx = wikipedia.summary(question, sentences=2)
+            context = wikipedia.summary(question, sentences=2)
         except Exception:
             pass
-        prompt = textwrap.dedent(f"""
-            You are a general AI assistant. I will ask you a question.
-            Report your thoughts briefly, then finish with the template:
-            FINAL ANSWER: <short precise answer>.
-            Question: {question}
-            Context: {ctx}
-            """).strip()
-        out = self.llm(prompt)[0]["generated_text"]
-        return out.split("FINAL ANSWER:")[-1].strip()
 
-    # ────────────────────────────────────────────────────────────────────
-    def __call__(self, q: str) -> str:
-        ql = q.lower()
+        prompt = textwrap.dedent(
+            f"""
+            You are an expert assistant. Answer in one short sentence or the
+            exact string/number requested.
 
-        # 1) albums
-        if m := self.PAT_ALBUMS.search(ql):
-            return albums_between(m[1].title(), int(m[2]), int(m[3]))
-
-        # 2) reverse-word puzzle
-        if ql.startswith(".rewsna"):
-            return reverse_left_puzzle(q)
+            Question: {question}
+            Context: {context}
+            Answer:
+            """
+        ).strip()
 
-        # 3) non-commutative subset
-        if "counter-examples" in ql and "|*" in q:
-            nc = non_comm_subset(q)
-            if nc:
-                return nc
+        reply   = self.llm(prompt)[0]["generated_text"]
+        answer  = reply.split("Answer:")[-1].strip().rstrip(".")
+        return answer or "I don't know"
 
-        # fallback LLM
-        return self._llm_answer(q)
 
 # ────────────────────────────────────────────────────────────────────────────
-#  run & submit
+#  RUN & SUBMIT
 # ────────────────────────────────────────────────────────────────────────────
 def run_and_submit_all(profile: gr.OAuthProfile | None):
     if not profile:
         return "Please login first.", None
-    username = profile.username
-    agent    = SmartAgent()
-
-    space_id   = os.getenv("SPACE_ID") or "local"
-    code_link  = f"https://huggingface.co/spaces/{space_id}/tree/main"
+    username  = profile.username
+    agent     = SmartAgent()
+    space_id  = os.getenv("SPACE_ID") or "local"
+    code_link = f"https://huggingface.co/spaces/{space_id}/tree/main"
 
     qs = requests.get(f"{API_URL}/questions", timeout=30).json()
 
-    rows, payload = [], []
+    rows, answers_payload = [], []
     for item in qs:
-        tid, q = item["task_id"], item["question"]
+        tid, q_text = item["task_id"], item["question"]
         try:
-            ans = agent(q)
+            ans = agent(q_text)
         except Exception as e:
             ans = f"AGENT ERROR: {e}"
 
-        # prepend FINAL ANSWER tag if GAIA format
-        final = f"FINAL ANSWER: {ans}" if GAIA_FORMAT else ans
-        field = "model_answer" if GAIA_FORMAT else "submitted_answer"
+        final = f"FINAL ANSWER: {ans}" if GAIA_FMT else ans
+        field = "model_answer" if GAIA_FMT else "submitted_answer"
 
-        payload.append({"task_id": tid, field: final})
-        rows.append({"Task ID": tid, "Question": q, "Answer": final})
+        answers_payload.append({"task_id": tid, field: final})
+        rows.append({"Task ID": tid, "Question": q_text, "Answer": final})
 
     submission = {
         "username": username,
         "agent_code": code_link,
-        "answers": payload,
+        "answers": answers_payload,
     }
 
-    resp = requests.post(f"{API_URL}/submit", json=submission, timeout=120)
-    resp.raise_for_status()
-    r = resp.json()
+    r = requests.post(f"{API_URL}/submit", json=submission, timeout=120).json()
 
-    score_line = (
+    mode   = "GAIA format" if GAIA_FMT else "Course format"
+    status = (
+        f"Submitted in **{mode}** – "
         f"Score: {r.get('score')} % "
         f"({r.get('correct_count')}/{r.get('total_attempted')})"
     )
-    mode = "GAIA format" if GAIA_FORMAT else "Course format"
-    status = f"Submitted in **{mode}** – {score_line}"
     return status, pd.DataFrame(rows)
 
+
 # ────────────────────────────────────────────────────────────────────────────
 #  UI
 # ────────────────────────────────────────────────────────────────────────────
 with gr.Blocks() as demo:
-    gr.Markdown("# GAIA Agents-Course – SmartAgent")
+    gr.Markdown("# GAIA Agents-Course – SmartAgent demo")
     gr.Markdown(
-        "Click the button to run the 20 validation questions, submit, "
-        "and display the score.<br>"
-        f"*Current output mode*: **{'GAIA' if GAIA_FORMAT else 'Course'}** "
-        "(toggle with the `GAIA_FORMAT` env-var)."
+        "Press the button to answer the 20 validation questions, submit, and see the score. "
+        f"*Current output mode*: **{'GAIA' if GAIA_FMT else 'Course'}** "
+        "(set via `GAIA_FORMAT` env-var)."
     )
+
     gr.LoginButton()
-    btn   = gr.Button("Run Evaluation & Submit")
-    stat  = gr.Markdown()
-    table = gr.DataFrame(wrap=True, interactive=False)
+    run_btn = gr.Button("Run Evaluation & Submit")
+    status  = gr.Markdown()
+    table   = gr.DataFrame(wrap=True, interactive=False)
 
-    btn.click(run_and_submit_all, outputs=[stat, table])
+    run_btn.click(run_and_submit_all, outputs=[status, table])
 
 if __name__ == "__main__":
     demo.launch()