train.py

# ================================================================
#  ASAD AI — BEST BRAIN TRAINER  v3.0
#  Loads 2 real HuggingFace datasets:
#    1. angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k  (38K rows)
#    2. TeichAI/DeepSeek-v4-Pro-Agent                    (4K rows)
#  Extracts Q&A pairs → trains 4-layer neural net
#  Auto-saves to /data/ (HF persistent storage)
#  Runs every 24h via background thread in app.py
# ================================================================

import os, json, re, time, datetime, logging, random
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from sklearn.preprocessing import LabelEncoder

logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s [TRAIN] %(message)s",
    datefmt="%H:%M:%S"
)
log = logging.getLogger(__name__)

# ── Paths ────────────────────────────────────────────────────────
STORAGE_DIR = os.environ.get("STORAGE_DIR", "/data")
os.makedirs(STORAGE_DIR, exist_ok=True)
MODEL_PATH  = os.path.join(STORAGE_DIR, "asad_ai_best.pth")
INFO_PATH   = os.path.join(STORAGE_DIR, "model_info.json")
DATA_PATH   = os.path.join(STORAGE_DIR, "training_data.json")
LOG_PATH    = os.path.join(STORAGE_DIR, "train_log.jsonl")

EPOCH_TIMEOUT = 25 * 60   # 25-min safety guard (well inside 30-min HF limit)

# ================================================================
#  BASE INTENT DATA  (always available — no network needed)
# ================================================================

BASE_DATA = {
    "intents": [
        {"tag": "greeting",
         "patterns": ["hello","hi","hey","assalamualaikum","salam","kya haal hai",
                      "kaise ho","good morning","good evening","namaste","howdy",
                      "hola","aadab","salam bhai","kya chal raha hai","how are you",
                      "what's up","whats up","sup","heyy","hiii"],
         "responses": ["Walaikum Assalam! 😊 Main Asad AI hoon — kya madad kar sakta hoon?",
                       "Hello! Bohot khushi hui milke 🤖 Batao kya chahiye?",
                       "Salam! Main aapki help ke liye ready hoon! 🚀",
                       "Hi there! Asad AI at your service! Kuch poochho!"]},

        {"tag": "goodbye",
         "patterns": ["bye","goodbye","alvida","phir milenge","khuda hafiz","allah hafiz",
                      "see you","take care","chal chalta hoon","jaa raha hoon","later",
                      "ttyl","bye bye","tata","farewell","good night","shab bakhair"],
         "responses": ["Allah Hafiz! 👋 Dobara aana!",
                       "Khuda Hafiz! Apna khayal rakhna! 🙏",
                       "Bye! Jab bhi zaroorat ho, main yahan hoon! 😊"]},

        {"tag": "identity",
         "patterns": ["tumhara naam kya hai","aap kaun ho","who are you","your name",
                      "naam batao","tum kya ho","introduce yourself","are you a robot",
                      "are you ai","kya tum ai ho","tell me about yourself"],
         "responses": ["Main Asad AI hoon! 🤖 Ek custom-trained bilingual chatbot — Urdu aur English dono!",
                       "Mera naam Asad AI hai — aapki madad ke liye 24/7 ready! 🧠",
                       "I am Asad AI — Pakistan ka smart AI assistant! 🇵🇰"]},

        {"tag": "thanks",
         "patterns": ["shukriya","thank you","thanks","jazakallah","meherbani",
                      "bahut acha","zabardast","great","wonderful","bohot khoob",
                      "mashallah","waah","perfect","excellent","amazing","awesome"],
         "responses": ["Koi baat nahi! 😊 Aur kuch chahiye?",
                       "Bohot khushi hui madad karke! 🤖",
                       "Always welcome! Koi bhi sawaal ho toh poochho!",
                       "Jazakallah Khair! 🌟"]},

        {"tag": "help",
         "patterns": ["help","madad","kya kar sakte ho","what can you do",
                      "features","capabilities","kya jaante ho","tum kya kya jaante ho",
                      "what do you know","guide karo"],
         "responses": ["Main kar sakta hoon:\n✅ Urdu+English conversation\n✅ Coding help\n✅ Math explain karna\n✅ Pakistan ke baare mein\n✅ Motivation\n✅ General knowledge\nPoochho! 🤖"]},

        {"tag": "joke",
         "patterns": ["joke sunao","funny baat batao","hasa do","tell me a joke",
                      "joke","funny","mazak","latifa","mujhe hasao"],
         "responses": ["Teacher: 2+2 kya hai?\nStudent: Depend karta hai programming language par! 😄",
                       "Why do programmers prefer dark mode?\nBecause light attracts bugs! 🐛😂",
                       "Ek AI ne doosra AI se kaha: 'Tu ache code likhta hai!'\nDoosra: 'Shukriya, tujhse seekha!' 🤖😂"]},

        {"tag": "pakistan",
         "patterns": ["pakistan","lahore","karachi","islamabad","pakistan ke baare mein",
                      "about pakistan","pakistan zindabad","mera watan","pakistani"],
         "responses": ["Pakistan ek pyara mulk hai! 🇵🇰\n📍 Capital: Islamabad\n🗣️ Urdu — qaumi zaban\n🏏 Cricket world champions!\nPakistan Zindabad! 💚",
                       "Pakistan 14 August 1947 ko azaad hua — Masha Allah! 🇵🇰"]},

        {"tag": "programming",
         "patterns": ["coding","programming","python","code","developer","software",
                      "machine learning","AI","web development","bug","error","debug",
                      "github","javascript","html","css","data science"],
         "responses": ["Python se shuru karo — sabse aasaan aur powerful! 🐍\nFreeCodeCamp, YouTube Urdu tutorials try karo!",
                       "AI/ML ke liye: Python + PyTorch + HuggingFace — yahi main use karta hoon! 🤖"]},

        {"tag": "motivation",
         "patterns": ["motivate karo","i am sad","mein udaas hoon","discouraged",
                      "give up","haar gaya","zindagi mushkil hai","inspire karo",
                      "motivational quote","himmat dou"],
         "responses": ["Iqbal ne kaha:\n'Sitaron se aage jahan aur bhi hain!'\nTu capable hai — bas chal! 💪🌟",
                       "Har failure ek lesson hai! Einstein bhi school mein fail hua tha! 🚀"]},

        {"tag": "math",
         "patterns": ["math","maths","mathematics","calculate","calculation","algebra",
                      "geometry","calculus","equation","formula","percentage","hisaab",
                      "numbers","statistics","probability","2+2","solve karo"],
         "responses": ["Math mein madad kar sakta hoon! Kaunsa sawaal hai? 📐",
                       "Equation share karo — main step by step explain karunga! 🧮"]},

        {"tag": "science",
         "patterns": ["science","physics","chemistry","biology","scientific","experiment",
                      "theory","atom","molecule","gravity","energy","force","light",
                      "evolution","dna","cells","planets","solar system"],
         "responses": ["Science bohot interesting hai! Kaunsa topic chahiye? 🔬",
                       "Physics, Chemistry ya Biology — batao kya poochna hai! ⚛️"]},

        {"tag": "history",
         "patterns": ["history","itihas","tarikh","historical","war","battle","empire",
                      "civilization","ancient","mughal","british raj","independence",
                      "world war","1947","partition"],
         "responses": ["History fascinating hai! Pakistan ki 1947 ki azaadi — ek ajeeb daastaan! 📜",
                       "Kaunse waqt ka history poochna hai? Main batata hoon! 🏛️"]},

        {"tag": "food",
         "patterns": ["khana","food","biryani","nihari","karahi","chai","tea","coffee",
                      "recipe","kya khayein","hungry","bhook","Pakistani food","dhaba"],
         "responses": ["Pakistani khana duniya ka best! 🍛\n⭐ Biryani — king!\n⭐ Nihari — soul food!\n⭐ Chai — life! ☕",
                       "Biryani: chawal + gosht + masale + dum = perfection! 😄🍚"]},

        {"tag": "general_knowledge",
         "patterns": ["duniya ki capital","world capital","largest","smallest","population",
                      "moon","sun","earth","space","interesting facts","did you know",
                      "gk","trivia","amazing facts","general knowledge"],
         "responses": ["Interesting facts:\n🌍 Russia — sabse bada mulk\n🏔️ K2 — Pakistan mein (2nd highest)\n🌊 Pacific — sabse bada ocean\nAur kuch poochho! 🧠"]},

        {"tag": "creator",
         "patterns": ["tumhe kisne banaya","who created you","creator kaun hai",
                      "asad kaun hai","who is asad","developer kaun hai","made by"],
         "responses": ["Mujhe Asad ne banaya! 👨‍💻🇵🇰 Ek Pakistani AI developer — mera ustaad!",
                       "Asad — mera creator, mera trainer! Unhone PyTorch se mujhe banaya! 🤖"]},

        {"tag": "unknown",
         "patterns": [],
         "responses": ["Maafi chahta hoon, samajh nahi aaya 🤔 Thoda aur detail mein poochho?",
                       "Interesting sawaal! Lekin abhi mujhe pata nahi — main seekh raha hoon! 😊",
                       "Sorry! Main abhi is topic par trained nahi hoon. Kuch aur poochho! 🤖"]}
    ]
}

# ================================================================
#  DATASET LOADER  — HuggingFace se Q&A pairs extract karo
# ================================================================

def load_hf_datasets(max_claude=600, max_deepseek=200):
    """
    Downloads both datasets and extracts (question, category) pairs
    to augment our intent classifier.
    Returns list of {"tag": str, "patterns": [str], "responses": [str]}
    """
    extra_intents = {}   # tag → {patterns, responses}

    # ── 1. Claude Opus reasoning dataset ────────────────────────
    try:
        log.info("📥 Loading claude-opus-4.6-4.7-reasoning-8.7k ...")
        from datasets import load_dataset
        ds_claude = load_dataset(
            "angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k",
            split="train",
            streaming=True   # streaming = no full disk cache needed
        )
        count = 0
        for row in ds_claude:
            if count >= max_claude:
                break
            category = (row.get("category") or "general").strip().lower()
            category = re.sub(r'[^a-z0-9_]', '_', category)
            messages  = row.get("messages", [])

            # Extract user question + assistant answer
            user_msg  = next((m["content"] for m in messages if m["role"] == "user"), None)
            asst_msg  = next((m["content"] for m in messages if m["role"] == "assistant"), None)
            if not user_msg or not asst_msg:
                continue

            # Clean: strip <think>...</think> blocks from assistant
            clean_asst = re.sub(r'<think>.*?</think>', '', asst_msg, flags=re.DOTALL).strip()
            if len(clean_asst) < 20 or len(user_msg) < 5:
                continue

            # Truncate for storage
            user_q  = user_msg[:200].strip()
            clean_a = clean_asst[:400].strip()

            tag = f"ds_claude_{category}"
            if tag not in extra_intents:
                extra_intents[tag] = {"tag": tag, "patterns": [], "responses": []}
            if len(extra_intents[tag]["patterns"]) < 40:
                extra_intents[tag]["patterns"].append(user_q)
            if len(extra_intents[tag]["responses"]) < 15:
                extra_intents[tag]["responses"].append(clean_a)
            count += 1

        log.info(f"✅ Claude dataset: {count} rows → {len([k for k in extra_intents if 'claude' in k])} intent categories")

    except Exception as e:
        log.warning(f"⚠️  Claude dataset load failed: {e} — using base data only")

    # ── 2. DeepSeek agent traces dataset ────────────────────────
    try:
        log.info("📥 Loading TeichAI/DeepSeek-v4-Pro-Agent ...")
        from datasets import load_dataset
        ds_deepseek = load_dataset(
            "TeichAI/DeepSeek-v4-Pro-Agent",
            split="train",
            streaming=True
        )
        count = 0
        for row in ds_deepseek:
            if count >= max_deepseek:
                break
            prompt = (row.get("prompt") or "").strip()
            if len(prompt) < 10:
                continue

            # Extract first assistant response from traces
            traces = row.get("traces", [])
            asst_response = None
            for t in traces:
                if isinstance(t, dict) and t.get("type") == "message":
                    msg = t.get("message", {})
                    if msg.get("role") == "assistant":
                        content = msg.get("content", [])
                        for c in content:
                            if isinstance(c, dict) and c.get("type") == "text":
                                txt = c.get("text", "").strip()
                                if len(txt) > 30:
                                    asst_response = txt[:400]
                                    break
                    if asst_response:
                        break

            if not asst_response:
                continue

            tag = "ds_deepseek_coding"
            if tag not in extra_intents:
                extra_intents[tag] = {"tag": tag, "patterns": [], "responses": []}
            if len(extra_intents[tag]["patterns"]) < 50:
                extra_intents[tag]["patterns"].append(prompt[:200])
            if len(extra_intents[tag]["responses"]) < 20:
                extra_intents[tag]["responses"].append(asst_response)
            count += 1

        log.info(f"✅ DeepSeek dataset: {count} rows → coding intent augmented")

    except Exception as e:
        log.warning(f"⚠️  DeepSeek dataset load failed: {e} — using base data only")

    # Filter: only keep intents with ≥3 patterns AND ≥1 response
    valid = [v for v in extra_intents.values()
             if len(v["patterns"]) >= 3 and len(v["responses"]) >= 1]
    log.info(f"📊 Extra intents from HF datasets: {len(valid)}")
    return valid


# ================================================================
#  MODEL
# ================================================================

class AsadAIModel(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(input_size, hidden_size),
            nn.LayerNorm(hidden_size),
            nn.GELU(),
            nn.Dropout(0.3),

            nn.Linear(hidden_size, hidden_size),
            nn.LayerNorm(hidden_size),
            nn.GELU(),
            nn.Dropout(0.2),

            nn.Linear(hidden_size, hidden_size // 2),
            nn.LayerNorm(hidden_size // 2),
            nn.GELU(),
            nn.Dropout(0.15),

            nn.Linear(hidden_size // 2, output_size)
        )

    def forward(self, x):
        return self.net(x)


class ChatDataset(Dataset):
    def __init__(self, X, y):
        self.X = torch.FloatTensor(X)
        self.y = torch.LongTensor(y)
    def __len__(self): return len(self.X)
    def __getitem__(self, i): return self.X[i], self.y[i]


# ================================================================
#  TEXT UTILS
# ================================================================

def clean(text):
    text = str(text).lower().strip()
    return re.sub(r'[^\w\s]', '', text)

def build_vocab(intents):
    vocab, pats, tags = set(), [], []
    for intent in intents:
        for p in intent["patterns"]:
            words = clean(p).split()
            vocab.update(words)
            pats.append(clean(p))
            tags.append(intent["tag"])
    return sorted(vocab), pats, tags

def bow(text, vocab):
    v = np.zeros(len(vocab), dtype=np.float32)
    for w in clean(text).split():
        if w in vocab:
            v[vocab.index(w)] = 1.0
    return v

def append_log(entry):
    try:
        with open(LOG_PATH, 'a', encoding='utf-8') as f:
            f.write(json.dumps(entry, ensure_ascii=False) + '\n')
    except Exception:
        pass


# ================================================================
#  MAIN TRAINING
# ================================================================

def run_training():
    """
    Full pipeline:
      1. Load HF datasets (streaming, no full cache)
      2. Merge with base intents
      3. Train 4-layer neural net
      4. Save model + metadata to /data/
    Returns (model, vocab, le, all_intents) or None on error.
    """
    start = time.time()
    ts    = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
    log.info(f"{'='*55}")
    log.info(f"🚀 Training started: {ts}")

    try:
        # ── Step 1: Build dataset ────────────────────────────────
        extra_intents = load_hf_datasets(max_claude=600, max_deepseek=200)
        all_intents   = BASE_DATA["intents"] + extra_intents
        merged_data   = {"intents": all_intents}

        # Save merged data snapshot
        with open(DATA_PATH, 'w', encoding='utf-8') as f:
            json.dump(merged_data, f, ensure_ascii=False, indent=2)

        vocab_list, all_pats, all_tags = build_vocab(all_intents)
        log.info(f"📊 Vocab: {len(vocab_list)} words | Patterns: {len(all_pats)} | Intents: {len(set(all_tags))}")

        if len(all_pats) < 10:
            log.error("Not enough training data!")
            return None

        le = LabelEncoder()
        le.fit(all_tags)

        X = np.array([bow(p, vocab_list) for p in all_pats])
        y = le.transform(all_tags)

        # ── Step 2: Model config ─────────────────────────────────
        IN     = len(vocab_list)
        H      = 256
        OUT    = len(le.classes_)
        EPOCHS = 400
        BATCH  = max(4, min(32, len(X) // 4))
        LR     = 0.001

        model     = AsadAIModel(IN, H, OUT)
        criterion = nn.CrossEntropyLoss(label_smoothing=0.1)
        optimizer = torch.optim.AdamW(model.parameters(), lr=LR, weight_decay=1e-4)
        scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=EPOCHS)

        ds     = ChatDataset(X, y)
        loader = DataLoader(ds, batch_size=BATCH, shuffle=True, drop_last=False)

        # ── Step 3: Train ────────────────────────────────────────
        best_loss = float('inf')
        best_acc  = 0.0

        for epoch in range(EPOCHS):
            # 25-min timeout guard
            if time.time() - start > EPOCH_TIMEOUT:
                log.warning("⚠️  25-min timeout — stopping early")
                break

            model.train()
            tot_loss, correct, total = 0, 0, 0
            for bx, by in loader:
                optimizer.zero_grad()
                out  = model(bx)
                loss = criterion(out, by)
                loss.backward()
                torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
                optimizer.step()
                tot_loss += loss.item()
                pred = out.argmax(1)
                correct += (pred == by).sum().item()
                total   += by.size(0)
            scheduler.step()

            avg_loss = tot_loss / len(loader)
            acc      = correct / total * 100

            if avg_loss < best_loss:
                best_loss = avg_loss
                best_acc  = acc
                torch.save(model.state_dict(), MODEL_PATH)

            if (epoch + 1) % 100 == 0:
                log.info(f"  Epoch {epoch+1:4d}/{EPOCHS} | Loss {avg_loss:.4f} | Acc {acc:.1f}%")

        # ── Step 4: Load best + save metadata ───────────────────
        model.load_state_dict(torch.load(MODEL_PATH, map_location='cpu', weights_only=True))
        model.eval()

        elapsed = round(time.time() - start, 1)
        info = {
            "vocab"         : vocab_list,
            "tags"          : list(le.classes_),
            "input_size"    : IN,
            "hidden_size"   : H,
            "output_size"   : OUT,
            "best_loss"     : round(best_loss, 5),
            "best_acc"      : round(best_acc, 2),
            "trained_at"    : ts,
            "elapsed_s"     : elapsed,
            "patterns_n"    : len(all_pats),
            "intents_n"     : len(set(all_tags)),
            "hf_extra_n"    : len(extra_intents),
        }
        with open(INFO_PATH, 'w', encoding='utf-8') as f:
            json.dump(info, f, ensure_ascii=False, indent=2)

        log.info(f"✅ Done in {elapsed}s | Loss={best_loss:.4f} | Acc={best_acc:.1f}% | Intents={OUT}")
        append_log({"event": "done", "ts": ts, "loss": best_loss,
                    "acc": best_acc, "elapsed_s": elapsed,
                    "intents": OUT, "patterns": len(all_pats)})

        return model, vocab_list, le, merged_data

    except Exception as e:
        log.error(f"❌ Training failed: {e}")
        append_log({"event": "error", "ts": ts, "error": str(e)})
        return None


# ── Standalone run ────────────────────────────────────────────────
#if __name__ == "__main__":
   # result = run_training()
    #if result:
        #log.info("✅ Model ready at /data/asad_ai_best.pth")
    #else:
        #log.error("❌ Training failed — check logs")