Upload scripts/augment-fill-gaps.py with huggingface_hub

scripts/augment-fill-gaps.py

@@ -0,0 +1,291 @@
+#!/usr/bin/env python3
+"""
+Gap-filling dataset augmentation.
+Targets the specific gap: 0.5B at 91.2% vs 1.5B at 48.2%.
+Root cause: 0.5B uses all 7 LoRA targets + pure tool-calling data.
+            1.5B uses only 4 targets + chat data dilution.
+
+This creates pure tool-calling data with no chit-chat to close the gap.
+All examples use ALL 7 linear modules as targets (matching the v2 config).
+"""
+import json, random
+from pathlib import Path
+
+random.seed(42)
+OUT = Path("gap-filled")
+OUT.mkdir(exist_ok=True)
+
+TOOLS = [
+    {"type": "function", "function": {"name": "get_weather", "description": "Get weather for a city", "parameters": {"type": "object", "properties": {"location": {"type": "string"}, "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}}, "required": ["location"]}}},
+    {"type": "function", "function": {"name": "get_time", "description": "Get time for a city", "parameters": {"type": "object", "properties": {"location": {"type": "string"}}, "required": ["location"]}}},
+    {"type": "function", "function": {"name": "search_web", "description": "Search the web for information", "parameters": {"type": "object", "properties": {"query": {"type": "string"}}, "required": ["query"]}}},
+    {"type": "function", "function": {"name": "calculator", "description": "Calculate a math expression", "parameters": {"type": "object", "properties": {"expression": {"type": "string"}}, "required": ["expression"]}}},
+    {"type": "function", "function": {"name": "get_stock_price", "description": "Get current stock price", "parameters": {"type": "object", "properties": {"ticker": {"type": "string"}}, "required": ["ticker"]}}},
+    {"type": "function", "function": {"name": "translate_text", "description": "Translate text to a language", "parameters": {"type": "object", "properties": {"text": {"type": "string"}, "target_lang": {"type": "string"}}, "required": ["text", "target_lang"]}}},
+    {"type": "function", "function": {"name": "book_flight", "description": "Book a flight between cities", "parameters": {"type": "object", "properties": {"origin": {"type": "string"}, "destination": {"type": "string"}, "date": {"type": "string"}}, "required": ["origin", "destination", "date"]}}},
+    {"type": "function", "function": {"name": "send_email", "description": "Send an email", "parameters": {"type": "object", "properties": {"to": {"type": "string"}, "subject": {"type": "string"}, "body": {"type": "string"}}, "required": ["to", "subject"]}}},
+    {"type": "function", "function": {"name": "get_news", "description": "Get news articles for a topic", "parameters": {"type": "object", "properties": {"topic": {"type": "string"}, "count": {"type": "integer"}}, "required": ["topic"]}}},
+    {"type": "function", "function": {"name": "get_restaurant_info", "description": "Get info about a restaurant", "parameters": {"type": "object", "properties": {"name": {"type": "string"}, "location": {"type": "string"}}, "required": ["name"]}}},
+    {"type": "function", "function": {"name": "set_reminder", "description": "Set a reminder for a time", "parameters": {"type": "object", "properties": {"text": {"type": "string"}, "time": {"type": "string"}}, "required": ["text", "time"]}}},
+    {"type": "function", "function": {"name": "get_directions", "description": "Get directions between locations", "parameters": {"type": "object", "properties": {"origin": {"type": "string"}, "destination": {"type": "string"}, "mode": {"type": "string", "enum": ["driving", "walking", "transit"]}}, "required": ["origin", "destination"]}}},
+]
+
+def m(role, content=None, tc=None):
+    x = {"role": role}
+    if content is not None: x["content"] = content
+    if tc: x["tool_calls"] = tc
+    return x
+
+def tc(name, args):
+    return [{"function": {"name": name, "arguments": json.dumps(args, ensure_ascii=False)}}]
+
+def subs(names):
+    return [t for t in TOOLS if t["function"]["name"] in names]
+
+def save(name, examples):
+    path = OUT / f"{name}.jsonl"
+    with open(path, "w") as f:
+        for ex in examples:
+            f.write(json.dumps(ex, ensure_ascii=False) + "\n")
+    print(f"  {name}.jsonl: {len(examples)} examples")
+
+all_examples = []
+
+# ── GAP 1: Pure tool-calling (no chat, no diluted examples) ─────────────
+# Root cause: 1.5B uses 48.2% because training included chat data.
+# Fix: examples that FORCE tool use — every query MUST result in a tool call.
+def gap1_pure_tc(n=100):
+    cities = ["Bangkok","Tokyo","London","Paris","New York","Dubai","Singapore","Berlin","Rome","Madrid",
+              "Seoul","Mumbai","Sydney","Toronto","Moscow","Istanbul","Amsterdam","Prague","Vienna","Oslo"]
+    tickers = ["AAPL","GOOGL","MSFT","TSLA","NVDA","AMD","AMZN","META","NFLX","SPOT"]
+    topics = ["AI","climate change","renewable energy","space exploration","quantum computing",
+              "cryptocurrency","electric vehicles","cybersecurity","biotechnology","robotics"]
+    langs = ["th","fr","de","es","it","pt","ja","ko","zh","ar"]
+    origins = ["BKK","NYC","LHR","CDG","NRT","DXB","SFO","LAX","HKG","SIN"]
+    dests = ["NRT","LAX","CDG","HKG","SIN","BKK","LHR","JFK","DXB","SFO"]
+    dates = ["2026-10-01","2026-11-15","2026-12-25","2027-01-10","2027-02-14","2027-03-20","2027-04-05","2027-05-01"]
+    exprs = ["2+2","15/100*200","sqrt(144)","sin(pi/2)","log(100)","3**3","(5+3)*2","100/3"]
+    emails = [("alice@co.com","Project update","Done"),("bob@firm.com","Meeting","3pm"),
+              ("carol@org.com","Report","Attached"),("dave@io.com","Question","Please review")]
+    rest_names = ["Sushi Bar","Pizza Place","Taco Stand","Noodle House"," Curry Shop"]
+    reminders = ["Buy groceries","Call mom","Doctor appointment","Team standup","Submit report"]
+    
+    examples = []
+    for i in range(n):
+        c = cities[i % len(cities)]
+        t = tickers[i % len(tickers)]
+        top = topics[i % len(topics)]
+        lg = langs[i % len(langs)]
+        o, d = origins[i % len(origins)], dests[(i+3) % len(dests)]
+        dt = dates[i % len(dates)]
+        ex = exprs[i % len(exprs)]
+        em = emails[i % len(emails)]
+        rn = rest_names[i % len(rest_names)]
+        rm = reminders[i % len(reminders)]
+        
+        # Cycle through tool types
+        tool_type = i % 10
+        if tool_type == 0:
+            examples.append({"messages": [m("user", f"Weather in {c}?"), m("assistant", tc=tc("get_weather", {"location": c, "unit": "celsius"}))], "tools": subs(["get_weather","get_time"])})
+        elif tool_type == 1:
+            examples.append({"messages": [m("user", f"Stock price of {t}"), m("assistant", tc=tc("get_stock_price", {"ticker": t}))], "tools": subs(["get_stock_price","get_news"])})
+        elif tool_type == 2:
+            examples.append({"messages": [m("user", f"Search for {top}"), m("assistant", tc=tc("search_web", {"query": top}))], "tools": subs(["search_web","get_news"])})
+        elif tool_type == 3:
+            examples.append({"messages": [m("user", f"Calculate {ex}"), m("assistant", tc=tc("calculator", {"expression": ex}))], "tools": subs(["calculator","get_stock_price"])})
+        elif tool_type == 4:
+            if o != d:
+                examples.append({"messages": [m("user", f"Book flight from {o} to {d} on {dt}"), m("assistant", tc=tc("book_flight", {"origin": o, "destination": d, "date": dt}))], "tools": subs(["book_flight","get_restaurant_info"])})
+        elif tool_type == 5:
+            examples.append({"messages": [m("user", f"Translate 'hello' to {lg}"), m("assistant", tc=tc("translate_text", {"text": "hello", "target_lang": lg}))], "tools": subs(["translate_text","search_web"])})
+        elif tool_type == 6:
+            examples.append({"messages": [m("user", f"News about {top}"), m("assistant", tc=tc("get_news", {"topic": top, "count": 5}))], "tools": subs(["get_news","search_web"])})
+        elif tool_type == 7:
+            examples.append({"messages": [m("user", f"Email {em[0]} about {em[1]}"), m("assistant", tc=tc("send_email", {"to": em[0], "subject": em[1], "body": em[2]}))], "tools": subs(["send_email","book_flight"])})
+        elif tool_type == 8:
+            examples.append({"messages": [m("user", f"Set reminder to {rm} at 3pm"), m("assistant", tc=tc("set_reminder", {"text": rm, "time": "15:00"}))], "tools": subs(["set_reminder","get_time","calculator"])})
+        elif tool_type == 9:
+            c2 = cities[(i+7) % len(cities)]
+            examples.append({"messages": [m("user", f"Directions from {c} to {c2}"), m("assistant", tc=tc("get_directions", {"origin": c, "destination": c2, "mode": random.choice(["driving","walking","transit"])}))], "tools": subs(["get_directions","get_weather","get_time"])})
+    return examples[:n]
+
+# ── GAP 2: Irrelevance (current only 60, proven at 93.3% accuracy) ──────
+# The 0.5B gets 93.3% irrelevance. More irrelevance data = better safety.
+def gap2_irrelevance(n=100):
+    queries = [
+        "Hello!","How are you?","What's your name?","Tell me a joke","Good morning!",
+        "Thanks for your help","Have a nice day","What is AI?","Explain quantum physics",
+        "Who won the world cup?","What is the capital of Thailand?","How tall is Everest?",
+        "What is the speed of light?","Who painted Starry Night?","What is DNA?",
+        "How do vaccines work?","What is climate change?","Explain gravity",
+        "What is the boiling point of water?","How many continents are there?",
+        "What is the largest desert?","Who invented the printing press?",
+        "What is the Fibonacci sequence?","How does photosynthesis work?",
+        "What is the meaning of life?","Tell me a fun fact","What is 2+2?",
+        "Who wrote The Great Gatsby?","What is the smallest country?",
+        "How deep is the ocean?","What causes earthquakes?","How do birds fly?",
+        "What is the human genome?","How does memory work?","What is consciousness?",
+        "Why is the sky blue?","What are black holes?","How old is the universe?",
+        "What is renewable energy?","How do solar panels work?","What is blockchain?",
+        "How does encryption work?","What is democracy?","What is art?",
+        "How are clouds formed?","What is the water cycle?","How do muscles grow?",
+        "What is nutrition?","How do languages evolve?","What is culture?",
+    ]
+    examples = []
+    for q in queries * (n // len(queries) + 1):
+        random.shuffle(TOOLS)
+        examples.append({"messages": [m("user", q), m("assistant", content=f"That's an interesting question. Let me answer directly.")], "tools": TOOLS[:random.randint(4, 8)]})
+    return examples[:n]
+
+# ── GAP 3: Arguments precision (45.7% → target 65%+) ────────────────────
+# The biggest accuracy gap. Create examples testing exact argument matching.
+def gap3_args_precision(n=80):
+    cities = ["Bangkok","Tokyo","London","Paris","New York","Dubai","Singapore","Berlin","Rome","Madrid"]
+    examples = []
+    for i in range(n):
+        c = cities[i % len(cities)]
+        examples.append({"messages": [m("user", f"Weather in {c}?"), m("assistant", tc=tc("get_weather", {"location": c, "unit": "celsius"}))], "tools": subs(["get_weather","get_time"])})
+        examples.append({"messages": [m("user", f"Weather in {c} in fahrenheit"), m("assistant", tc=tc("get_weather", {"location": c, "unit": "fahrenheit"}))], "tools": subs(["get_weather","get_time"])})
+        t = ["AAPL","GOOGL","MSFT","TSLA","NVDA"][i % 5]
+        examples.append({"messages": [m("user", f"Stock for {t}"), m("assistant", tc=tc("get_stock_price", {"ticker": t}))], "tools": subs(["get_stock_price","get_news","calculator"])})
+        lg = ["th","fr","de","es","it","ja"][i % 6]
+        examples.append({"messages": [m("user", f"Translate 'friend' to {lg}"), m("assistant", tc=tc("translate_text", {"text": "friend", "target_lang": lg}))], "tools": subs(["translate_text","search_web"])})
+    return examples[:n]
+
+# ── GAP 4: Parallel calls (multiple simultaneous tools) ─────────────────
+def gap4_parallel(n=60):
+    cities = ["Bangkok","Tokyo","London","Paris","New York","Dubai","Singapore","Berlin","Rome","Madrid"]
+    tickers = ["AAPL","GOOGL","MSFT","TSLA","NVDA"]
+    topics = ["AI","climate","sports","tech","health"]
+    examples = []
+    for i in range(n):
+        c1, c2 = cities[i % len(cities)], cities[(i+3) % len(cities)]
+        examples.append({"messages": [m("user", f"Weather in {c1} and {c2}"),
+            m("assistant", tc=tc("get_weather", {"location": c1}) + tc("get_weather", {"location": c2}))],
+            "tools": subs(["get_weather","get_time"])})
+        t1, t2 = tickers[i % len(tickers)], tickers[(i+1) % len(tickers)]
+        examples.append({"messages": [m("user", f"Stocks for {t1} and {t2}"),
+            m("assistant", tc=tc("get_stock_price", {"ticker": t1}) + tc("get_stock_price", {"ticker": t2}))],
+            "tools": subs(["get_stock_price","get_news"])})
+        c3 = cities[(i+5) % len(cities)]
+        top = topics[i % len(topics)]
+        examples.append({"messages": [m("user", f"Weather in {c3} and news about {top}"),
+            m("assistant", tc=tc("get_weather", {"location": c3}) + tc("get_news", {"topic": top, "count": 3}))],
+            "tools": subs(["get_weather","get_news","search_web","get_time"])})
+    return examples[:n]
+
+# ── GAP 5: Selection accuracy (hard negatives) ──────────────────────────
+def gap5_selection(n=60):
+    pairs = [
+        ("get_weather", "get_time"), ("get_stock_price", "get_news"),
+        ("search_web", "get_news"), ("book_flight", "get_restaurant_info"),
+        ("translate_text", "search_web"), ("send_email", "book_flight"),
+        ("calculator", "get_stock_price"), ("set_reminder", "get_time"),
+        ("get_directions", "book_flight"), ("get_weather", "get_restaurant_info"),
+    ]
+    a = {"get_weather": {"location": "Paris", "unit": "celsius"}, "get_time": {"location": "Paris"},
+         "get_stock_price": {"ticker": "AAPL"}, "get_news": {"topic": "latest", "count": 3},
+         "search_web": {"query": "latest news"}, "book_flight": {"origin": "BKK", "destination": "NRT", "date": "2026-09-01"},
+         "get_restaurant_info": {"name": "Sushi Bar"}, "translate_text": {"text": "hello", "target_lang": "th"},
+         "send_email": {"to": "a@b.com", "subject": "Hi", "body": "Hello"}, "calculator": {"expression": "2+2"},
+         "set_reminder": {"text": "test", "time": "12:00"}, "get_directions": {"origin": "A", "destination": "B"}}
+    examples = []
+    for correct, wrong in pairs * (n // len(pairs) + 1):
+        q = f"Need {correct}"
+        examples.append({"messages": [m("user", q), m("assistant", tc=tc(correct, a[correct]))],
+                         "tools": subs([correct, wrong])})
+    return examples[:n]
+
+# ── GAP 6: Multi-hop chains ─────────────────────────────────────────────
+def gap6_multihop(n=40):
+    cities = ["Bangkok","Tokyo","London","Paris"]
+    tickers = ["AAPL","GOOGL","MSFT","TSLA"]
+    examples = []
+    for i in range(n):
+        c = cities[i % len(cities)]
+        t = tickers[i % len(tickers)]
+        # weather → translate
+        examples.append({"messages": [m("user", f"Weather in {c}, then translate to Thai"),
+            m("assistant", tc=tc("get_weather", {"location": c}) + tc("translate_text", {"text": "result", "target_lang": "th"}))],
+            "tools": subs(["get_weather","translate_text","search_web"])})
+        # stock → search for analysis
+        examples.append({"messages": [m("user", f"Stock of {t} and find analyst opinions"),
+            m("assistant", tc=tc("get_stock_price", {"ticker": t}) + tc("search_web", {"query": f"{t} analyst opinion 2026"}))],
+            "tools": subs(["get_stock_price","search_web","get_news"])})
+        # book flight → check weather at destination
+        c2 = cities[(i+2) % len(cities)]
+        examples.append({"messages": [m("user", f"Book to {c} from BKK and check weather there"),
+            m("assistant", tc=tc("book_flight", {"origin": "BKK", "destination": c, "date": "2026-10-01"}) + tc("get_weather", {"location": c}))],
+            "tools": subs(["book_flight","get_weather","get_restaurant_info"])})
+    return examples[:n]
+
+# ── GAP 7: Multi-turn tracking ──────────────────────────────────────────
+def gap7_multiturn(n=40):
+    cities = ["Rome","Paris","Tokyo","London","Berlin","Madrid","Dubai","Seoul","Bangkok","Mumbai"]
+    tickers = ["NVDA","AMD","AAPL","MSFT","GOOGL","TSLA","AMZN","META"]
+    examples = []
+    for i in range(n):
+        c1, c2 = cities[i % len(cities)], cities[(i+1) % len(cities)]
+        t1, t2 = tickers[i % len(tickers)], tickers[(i+1) % len(tickers)]
+        # Weather → follow-up
+        examples.append({"messages": [
+            m("user", f"Weather in {c1}?"), m("assistant", tc=tc("get_weather", {"location": c1})),
+            m("tool", f"22C in {c1}"), m("user", f"And in {c2}?")],
+            "tools": subs(["get_weather","get_time"])})
+        # Stock → follow-up
+        examples.append({"messages": [
+            m("user", f"Price of {t1}?"), m("assistant", tc=tc("get_stock_price", {"ticker": t1})),
+            m("tool", "$800"), m("user", f"What about {t2}?")],
+            "tools": subs(["get_stock_price","get_news"])})
+        # News → search follow-up
+        examples.append({"messages": [
+            m("user", "Latest AI news"), m("assistant", tc=tc("get_news", {"topic": "AI", "count": 3})),
+            m("tool", "Story 1... Story 2..."), m("user", "Tell me more about story 1")],
+            "tools": subs(["get_news","search_web"])})
+    return examples[:n]
+
+# ── Generate all ────────────────────────────────────────────────────────
+GENERATORS = [
+    ("01-pure-tc", gap1_pure_tc, 100, "Pure tool-calling (no chat dilution)"),
+    ("02-irrelevance", gap2_irrelevance, 100, "Irrelevance expansion (proven 93.3%)"),
+    ("03-args-precision", gap3_args_precision, 80, "Argument precision (45.7% gap)"),
+    ("04-parallel", gap4_parallel, 60, "Parallel calls (multi-tool)"),
+    ("05-selection", gap5_selection, 60, "Selection hard negatives"),
+    ("06-multihop", gap6_multihop, 40, "Multi-hop chains"),
+    ("07-multiturn", gap7_multiturn, 40, "Multi-turn tracking"),
+]
+
+total = 0
+print("Gap-Filling Dataset Augmentation")
+print(f"Target: Close 0.5B(91.2%) → 1.5B(48.2%) gap")
+print("=" * 50)
+for name, gen_fn, count, desc in GENERATORS:
+    batch = gen_fn(count)
+    save(name, batch)
+    all_examples.extend(batch)
+    total += len(batch)
+
+combined = OUT / "all-gap-filled.jsonl"
+with open(combined, "w") as f:
+    for ex in all_examples:
+        f.write(json.dumps(ex, ensure_ascii=False) + "\n")
+
+print(f"\n{'=' * 50}")
+print(f"TOTAL: {total} gap-filling examples")
+print(f"{'=' * 50}")
+
+# ── Full training data composition ──────────────────────────────────────
+print(f"\n📊 Complete training dataset composition:")
+print(f"   v7 original:     2,424")
+print(f"   + 10x augment:     660")
+print(f"   + benchmark targ:  232")
+print(f"   + gap-fill:        {total}")
+print(f"   = TOTAL:          {2424 + 660 + 232 + total}")
+print(f"\n📈 Gap analysis addressed:")
+print(f"   Root cause 1 (chat dilution): 100 pure TC examples")
+print(f"   Root cause 2 (args 45.7%):    80 precision examples")
+print(f"   Root cause 3 (irrelevance):   100 more irrelevance (was 60)")
+print(f"   Root cause 4 (parallel):      60 multi-tool examples")
+print(f"   Root cause 5 (selection):     60 hard negative pairs")
+print(f"   Root cause 6 (multi-hop):     40 chain examples")
+print(f"   Root cause 7 (multi-turn):    40 context tracking")