Update training/train_agent.py

training/train_agent.py

@@ -1,301 +1,372 @@
 """
 training/train_agent.py — SQL Database Engineer Agent
-Unsloth + GRPO training script.
-Run on venue GPU (April 25-26) with compute credits.
-ENV_URL points to live HF Space environment.
+FIXED: Uses local DatabaseSimulator for rewards (no HF Space calls)
+- No shared singleton state
+- Real delta rewards (0.0 for wrong actions, 40-75pts for correct)
+- Clear reward difference teaches model to prefer create_index over inspect_query
 """
 
-import os
-import re
-import json
-import requests
-from datasets import Dataset
+import os, json, sys, time
+from pathlib import Path
 
-# ── Try importing Unsloth (GPU only) ─────────────────────────
+# ── GPU check ─────────────────────────────────────────────────
+UNSLOTH_AVAILABLE = False
 try:
+    import torch
+    if not torch.cuda.is_available():
+        print("❌ No GPU. Unsloth requires CUDA GPU.")
+        sys.exit(1)
     from unsloth import FastLanguageModel
     from trl import GRPOTrainer, GRPOConfig
+    from datasets import Dataset
     UNSLOTH_AVAILABLE = True
-except ImportError:
-    UNSLOTH_AVAILABLE = False
-    print("⚠️  Unsloth not available. Run: pip install unsloth trl")
+    print(f"✅ GPU: {torch.cuda.get_device_name(0)}")
+    print(f"✅ VRAM: {torch.cuda.get_device_properties(0).total_memory/1e9:.1f}GB")
+except ImportError as e:
+    print(f"❌ {e}\nRun: pip install unsloth trl transformers datasets accelerate")
+    sys.exit(1)
 
-# ─────────────────────────────────────────────
-#  CONFIG
-# ─────────────────────────────────────────────
+# Add project root so we can import DatabaseSimulator
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+from env.db_simulator import DatabaseSimulator
 
+# ── Config ────────────────────────────────────────────────────
 ENV_URL    = os.getenv("ENV_URL",    "https://junaid0600-sql-db-engineer-agent.hf.space")
 HF_TOKEN   = os.getenv("HF_TOKEN",  "")
-MODEL_NAME = os.getenv("MODEL_NAME", "unsloth/Qwen2.5-7B-Instruct")
+MODEL_NAME = os.getenv("MODEL_NAME", "unsloth/Qwen2.5-1.5B-Instruct")
 OUTPUT_DIR = os.getenv("OUTPUT_DIR", "./sdea-trained")
+MAX_STEPS  = int(os.getenv("MAX_STEPS", "100"))
 
-SYSTEM_PROMPT = """You are a senior database engineer. 
-Given the current database state with slow queries, choose the BEST action to improve performance.
-Think step by step:
-1. If you haven't inspected queries yet → use inspect_query
-2. If you haven't analyzed indexes → use analyze_indexes  
-3. If you know which index is missing → use create_index
-4. If query can be rewritten better → use rewrite_query
-5. If table is huge (1M+ rows) → use partition_table
-6. When performance target is reached → use submit_report
+print(f"\n[CONFIG] Model:     {MODEL_NAME}")
+print(f"[CONFIG] Max steps: {MAX_STEPS}")
+print(f"[CONFIG] Output:    {OUTPUT_DIR}\n")
 
-Respond with JSON only — no explanation, no markdown:
-{"action_type": "...", "payload": {...}}"""
+# ── System prompt ─────────────────────────────────────────────
+SYSTEM_PROMPT = """You are a senior database engineer fixing slow database queries.
+You will see slow queries and table structures. Choose the BEST action.
 
+Key insight: create_index on the RIGHT columns fixes slow queries.
+Wrong columns = no improvement. Right columns = massive improvement.
 
-# ─────────────────────────────────────────────
-#  REWARD FUNCTION (calls live HF Space)
-# ─────────────────────────────────────────────
+Respond with ONLY valid JSON:
+{"action_type": "create_index", "payload": {"table": "TABLE_NAME", "columns": ["COL1", "COL2"]}}
 
-def parse_action(text: str) -> dict:
-    """Safely parse model output into an action dict.
-
-    Supports markdown ```json blocks and returns a fallback action on failure.
-    """
-    if not isinstance(text, str):
-        text = str(text)
+Available actions: inspect_query, analyze_indexes, create_index, rewrite_query, analyze_statistics, submit_report"""
 
-    candidate = text.strip()
+# ── Load all 15 scenarios ─────────────────────────────────────
+def load_all_scenarios() -> list:
+    scenarios = []
+    base = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+    for fname in ["easy_scenarios.json", "medium_scenarios.json", "hard_scenarios.json"]:
+        path = os.path.join(base, "dataset", fname)
+        try:
+            with open(path) as f:
+                data = json.load(f)
+                scenarios.extend(data)
+                print(f"  ✅ Loaded {len(data)} from {fname}")
+        except FileNotFoundError:
+            print(f"  ⚠️ {fname} not found")
+    print(f"  Total: {len(scenarios)} scenarios\n")
+    return scenarios
 
-    fenced = re.search(r"```(?:json)?\s*(\{.*\})\s*```", candidate, re.S | re.I)
-    if fenced:
-        candidate = fenced.group(1).strip()
-    elif candidate.startswith("```") and candidate.endswith("```"):
-        candidate = candidate[3:-3].strip()
+ALL_SCENARIOS = load_all_scenarios()
 
+# ── Parse LLM output ─────────────────────────────────────────
+def parse_action(text: str) -> dict:
+    """Parse LLM output into action dict."""
     try:
-        action = json.loads(candidate)
-        if isinstance(action, dict):
-            action["_parse_fallback"] = False
-            return action
-    except json.JSONDecodeError:
+        text = text.strip()
+        for marker in ["```json", "```"]:
+            if marker in text:
+                parts = text.split(marker)
+                text = parts[1] if len(parts) > 1 else parts[0]
+        text = text.strip()
+        data = json.loads(text)
+        if "action_type" in data and "payload" in data:
+            return data
+    except Exception:
         pass
+    return None  # None = invalid JSON = penalized
 
-    start = candidate.find("{")
-    end = candidate.rfind("}")
-    if start != -1 and end != -1 and end > start:
-        try:
-            action = json.loads(candidate[start:end + 1])
-            if isinstance(action, dict):
-                action["_parse_fallback"] = False
-                return action
-        except json.JSONDecodeError:
-            pass
-
-    return {
-        "action_type": "inspect_query",
-        "payload": {"query_id": "q1"},
-        "_parse_fallback": True,
+
+# ── LOCAL reward function using DatabaseSimulator ─────────────
+def compute_local_reward(action: dict, scenario: dict) -> tuple:
+    """
+    Compute reward LOCALLY using DatabaseSimulator.
+    No HF Space calls. No shared state. Clean every time.
+    
+    Returns (reward_score, db_delta, milestone_bonus)
+    """
+    sim      = DatabaseSimulator(scenario)
+    baseline = sim.get_performance_score()
+    hints    = scenario.get("missing_index_hints", [])
+
+    action_type = action.get("action_type", "")
+    payload     = action.get("payload", {})
+
+    # Apply action to simulator
+    if action_type == "create_index":
+        result  = sim.apply_action("create_index", payload)
+        delta   = result.get("delta", 0.0)
+
+    elif action_type == "inspect_query":
+        # Investigation — small reward, no DB change
+        delta = 0.0
+
+    elif action_type == "analyze_indexes":
+        delta = 0.0
+
+    elif action_type == "rewrite_query":
+        result = sim.apply_action("rewrite_query", payload)
+        delta  = result.get("delta", 0.0)
+
+    elif action_type == "analyze_statistics":
+        result = sim.apply_action("analyze_statistics", payload)
+        delta  = result.get("delta", 0.0)
+
+    elif action_type == "partition_table":
+        result = sim.apply_action("partition_table", payload)
+        delta  = result.get("delta", 0.0)
+
+    elif action_type == "submit_report":
+        # Terminal: score based on how much DB improved so far
+        final = sim.get_performance_score()
+        improvement = max(0, final - baseline)
+        delta = improvement
+
+    else:
+        delta = -5.0  # Unknown action = penalty
+
+    final_score  = sim.get_performance_score()
+    improvement  = max(0.0, final_score - baseline)
+    max_possible = max(1.0, 100.0 - baseline)
+
+    # ── Reward components ─────────────────────────────────────
+    # 1. Step reward — different per action type
+    step_rewards = {
+        "inspect_query":     0.10,
+        "analyze_indexes":   0.10,
+        "create_index":      0.15,
+        "rewrite_query":     0.20,
+        "analyze_statistics":0.08,
+        "partition_table":   0.15,
+        "submit_report":     0.05,
     }
+    step_r = step_rewards.get(action_type, 0.001)
 
+    # 2. Delta reward — proportional to actual improvement
+    delta_r = min(0.70, (improvement / max_possible) * 0.70)
 
-def compute_reward(action_dict: dict, obs: dict, env_response: dict) -> float:
-    """Apply shaping rules on top of environment reward.
+    # 3. Milestone bonus — one-time for big improvements
+    milestone_r = 0.0
+    if improvement / max_possible >= 0.75:
+        milestone_r = 0.40
+    elif improvement / max_possible >= 0.50:
+        milestone_r = 0.25
+    elif improvement / max_possible >= 0.25:
+        milestone_r = 0.15
 
-    This strengthens the signal, penalizes repeated inspection, empty payloads,
-    and bounds final reward between 0.0 and 1.0.
-    """
-    base_score = env_response.get("score", 0.0)
-    try:
-        base_score = float(base_score)
-    except (TypeError, ValueError):
-        base_score = 0.0
-
-    if action_dict.get("_parse_fallback", False):
-        return 0.0
-
-    action_type = action_dict.get("action_type", "")
-    payload = action_dict.get("payload") or {}
-
-    if action_type == "inspect_query":
-        query_id = payload.get("query_id")
-        inspected_queries = []
-        if isinstance(obs, dict):
-            inspected_queries = obs.get("inspected_queries") or obs.get("inspected_query_ids") or []
-
-        if isinstance(inspected_queries, list) and query_id in inspected_queries:
-            return max(0.0, base_score - 0.2)
-
-    if not payload or (isinstance(payload, dict) and len(payload) == 0):
-        return max(0.0, base_score - 0.3)
-
-    valid_actions = {
-        "inspect_query",
-        "analyze_indexes",
-        "create_index",
-        "fix_query",
-        "optimize_query",
-    }
-    if action_type not in valid_actions:
-        return 0.0
+    # 4. Penalty for wrong index (delta=0 on create_index)
+    wrong_index_pen = 0.0
+    if action_type == "create_index" and delta <= 0.0:
+        wrong_index_pen = -0.15  # created useless index
 
-    if base_score > 0.9:
-        return min(1.0, base_score + 0.1)
+    total = step_r + delta_r + milestone_r + wrong_index_pen
+    total = max(0.001, min(0.999, total))
 
-    return base_score
+    return total, improvement, milestone_r
 
 
+# ── GRPO reward function ──────────────────────────────────────
 def reward_fn(prompts, completions, **kwargs):
     """
-    GRPO reward function — calls /step on live environment.
-    Returns list of float rewards, one per completion.
+    LOCAL reward — no HTTP calls, no shared state.
+    Each completion gets its own fresh DatabaseSimulator.
+    
+    Reward differences:
+      inspect_query (always):     0.10 + 0.0  = 0.10
+      create_index (wrong col):   0.15 - 0.15 = 0.001
+      create_index (right col):   0.15 + 0.60 = 0.75+
+    
+    GRPO will learn: right create_index >> inspect_query >> wrong create_index
     """
     rewards = []
-    task_ids = kwargs.get("task_ids", ["easy_s001"] * len(prompts))
 
     for i, (prompt, completion) in enumerate(zip(prompts, completions)):
         try:
-            text = completion[0]["content"] if isinstance(completion, list) else str(completion)
+            # Get text
+            if isinstance(completion, list):
+                text = completion[0].get("content", "") if completion else ""
+            else:
+                text = str(completion)
+
+            # Pick scenario (rotate through all)
+            scenario = ALL_SCENARIOS[i % len(ALL_SCENARIOS)]
+
+            # Parse action
             action = parse_action(text)
 
-            task_id = task_ids[i] if i < len(task_ids) else "easy_s001"
-            requests.post(
-                f"{ENV_URL}/reset",
-                json={"task_id": task_id},
-                timeout=15,
-            )
-
-            resp = requests.post(
-                f"{ENV_URL}/step",
-                json=action,
-                timeout=15,
-            )
-            env_response = resp.json()
-            score = compute_reward(
-                action_dict=action,
-                obs=env_response.get("observation", {}),
-                env_response=env_response,
-            )
-            score = max(0.0, min(1.0, float(score)))
+            if action is None:
+                # Invalid JSON output — penalize
+                rewards.append(0.001)
+                print(f"  [REWARD] scenario={scenario['id']} | "
+                      f"INVALID JSON | score=0.001")
+                continue
+
+            # Compute reward locally
+            score, improvement, milestone = compute_local_reward(action, scenario)
             rewards.append(score)
 
+            print(f"  [REWARD] scenario={scenario['id']} | "
+                  f"action={action.get('action_type')} | "
+                  f"improvement=+{improvement:.1f}pts | "
+                  f"milestone=+{milestone:.2f} | "
+                  f"score={score:.3f}")
+
         except Exception as e:
-            print(f"Reward fn error: {e}")
-            rewards.append(0.0)
+            print(f"  [REWARD] Error: {e}")
+            rewards.append(0.001)
 
     return rewards
 
 
-# ─────────────────────────────────────────────
-#  BUILD TRAINING DATASET
-# ─────────────────────────────────────────────
+# ── Build dataset ─────────────────────────────────────────────
+def build_dataset() -> Dataset:
+    examples = []
+    for i, s in enumerate(ALL_SCENARIOS):
+        tables_str  = json.dumps(s.get("tables", []))
+        queries_str = json.dumps(s.get("slow_queries", []))
+        hints_str   = json.dumps(s.get("missing_index_hints", []))
+
+        prompt = (
+            f"{SYSTEM_PROMPT}\n\n"
+            f"=== DATABASE STATE ===\n"
+            f"Scenario: {s['id']}\n"
+            f"Description: {s.get('description','')}\n"
+            f"Tables: {tables_str}\n"
+            f"Slow Queries: {queries_str}\n"
+            f"Missing Index Hints: {hints_str}\n"
+            f"Performance: {s.get('performance_score_baseline',0)}/100 "
+            f"→ Target: {s.get('target_score',85)}/100\n\n"
+            f"What action should you take? Output JSON only:"
+        )
+        examples.append({
+            "prompt":      prompt,
+            "scenario_id": s["id"],
+        })
 
-def build_dataset():
-    """Build training examples from all 15 Round 2 scenarios."""
-    scenarios = []
+    print(f"  ✅ Dataset: {len(examples)} examples")
+    return Dataset.from_list(examples)
 
-    # Load all scenario files
-    for fname in ["dataset/easy_scenarios.json",
-                  "dataset/medium_scenarios.json",
-                  "dataset/hard_scenarios.json"]:
-        try:
-            with open(fname) as f:
-                scenarios.extend(json.load(f))
-        except FileNotFoundError:
-            print(f"{fname} not found, skipping")
 
-    if not scenarios:
-        # Fallback: fetch from live environment
-        resp = requests.get(f"{ENV_URL}/tasks", timeout=15)
-        tasks = resp.json().get("tasks", [])
-        scenarios = [{"id": t["id"], "description": t["description"]} for t in tasks]
+# ── Generate plots ────────────────────────────────────────────
+def generate_plots(trainer):
+    import matplotlib
+    matplotlib.use("Agg")
+    import matplotlib.pyplot as plt
 
-    examples = []
-    for s in scenarios:
-        prompt = f"""{SYSTEM_PROMPT}
+    logs = [l for l in trainer.state.log_history if "loss" in l]
+    if not logs:
+        print("⚠️ No logs for plotting")
+        return
 
-Current Database State:
-- Scenario: {s.get('id', 'unknown')}
-- Description: {s.get('description', '')}
-- Tables: {json.dumps(s.get('tables', []))}
-- Slow Queries: {json.dumps(s.get('slow_queries', []))}
-- Performance Score: {s.get('performance_score_baseline', 0)} / 100
-- Target Score: {s.get('target_score', 85)}
+    steps  = [l.get("step", i) for i,l in enumerate(logs)]
+    losses = [l.get("loss", 0) for l in logs]
 
-What is your next action?"""
+    fig, ax = plt.subplots(1, 1, figsize=(8, 4))
+    fig.suptitle("GRPO Training — SQL Database Engineer Agent",
+                 fontsize=13, fontweight="bold")
 
-        examples.append({
-            "prompt":   prompt,
-            "task_id":  s.get("id", "easy_s001"),
-        })
+    ax.plot(steps, losses, "b-o", lw=2, ms=4)
+    ax.set_xlabel("Training Step")
+    ax.set_ylabel("Loss")
+    ax.set_title("Training Loss (↓ = model learning DBA pattern)")
+    ax.grid(True, alpha=0.3)
 
-    print(f"Built {len(examples)} training examples")
-    return Dataset.from_list(examples)
+    if losses:
+        ax.annotate(f"Start: {losses[0]:.4f}",
+                   xy=(steps[0], losses[0]),
+                   xytext=(steps[0]+1, losses[0]*1.1),
+                   fontsize=9, color="red")
+        ax.annotate(f"End: {losses[-1]:.4f}",
+                   xy=(steps[-1], losses[-1]),
+                   xytext=(steps[-1]-5, losses[-1]*1.1),
+                   fontsize=9, color="green")
 
+    plt.tight_layout()
+    plt.savefig("loss_curve.png", dpi=150, bbox_inches="tight")
+    print("✅ loss_curve.png saved")
+    print(f"   Loss: {losses[0]:.4f} → {losses[-1]:.4f}")
 
-# ─────────────────────────────────────────────
-#  MAIN TRAINING
-# ─────────────────────────────────────────────
 
+# ── Main ──────────────────────────────────────────────────────
 def train():
-    if not UNSLOTH_AVAILABLE:
-        print("Cannot train — Unsloth not installed")
-        print("Run: pip install unsloth trl transformers datasets accelerate")
-        return
-
-    print(f"🚀 Loading model: {MODEL_NAME}")
-    print(f"🌐 Environment: {ENV_URL}")
+    if not ALL_SCENARIOS:
+        print("❌ No scenarios found. Check dataset/ folder.")
+        sys.exit(1)
 
-    # Load model with Unsloth 4-bit quantization
+    print(f"⏳ Loading {MODEL_NAME}...")
     model, tokenizer = FastLanguageModel.from_pretrained(
         model_name     = MODEL_NAME,
-        max_seq_length = 4096,
+        max_seq_length = 2048,
         load_in_4bit   = True,
         token          = HF_TOKEN or None,
     )
-
-    # Add LoRA adapters
     model = FastLanguageModel.get_peft_model(
         model,
-        r              = 16,
-        lora_alpha     = 16,
-        target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
-                          "gate_proj", "up_proj", "down_proj"],
-        lora_dropout   = 0,
-        bias           = "none",
-        use_gradient_checkpointing = "unsloth",
+        r=16, lora_alpha=16,
+        target_modules=["q_proj","k_proj","v_proj","o_proj",
+                        "gate_proj","up_proj","down_proj"],
+        lora_dropout=0, bias="none",
+        use_gradient_checkpointing="unsloth",
+        random_state=42,
     )
+    print("✅ Model ready\n")
 
-    # Build dataset
     dataset = build_dataset()
 
-    # GRPO config
     config = GRPOConfig(
         output_dir                  = OUTPUT_DIR,
-        num_train_epochs            = 3,
-        per_device_train_batch_size = 2,
-        gradient_accumulation_steps = 8,
-        learning_rate               = 5e-5,
-        max_completion_length       = 256,
-        num_generations             = 4,
-        logging_steps               = 10,
-        save_steps                  = 50,
+        max_steps                   = MAX_STEPS,
+        per_device_train_batch_size = 1,
+        gradient_accumulation_steps = 4,
+        learning_rate               = 5e-6,
+        max_completion_length       = 150,
+        num_generations             = 4,   # compare 4 actions per step
+        temperature                 = 1.0,
+        logging_steps               = 1,
+        save_steps                  = 25,
+        save_total_limit            = 3,
         warmup_ratio                = 0.1,
         report_to                   = "none",
+        remove_unused_columns       = False,
     )
 
-    # Reward function wrapper
-    def reward_wrapper(prompts, completions, **kwargs):
-        task_ids = [ex.get("task_id", "easy_s001") for ex in kwargs.get("batch", [])]
-        return reward_fn(prompts, completions, task_ids=task_ids)
-
-    # Train
     trainer = GRPOTrainer(
-        model        = model,
-        tokenizer    = tokenizer,
-        reward_funcs = reward_wrapper,
-        args         = config,
+        model         = model,
+        tokenizer     = tokenizer,
+        reward_funcs  = reward_fn,
+        args          = config,
         train_dataset = dataset,
     )
 
-    print("🏋️  Starting GRPO training...")
+    print(f"🏋️  GRPO training — {MAX_STEPS} steps")
+    print("Watch for: improvement > 0 and score > 0.5 on create_index\n")
     trainer.train()
+    print("\n✅ Training complete!")
 
-    # Save
+    Path(OUTPUT_DIR).mkdir(parents=True, exist_ok=True)
     model.save_pretrained(f"{OUTPUT_DIR}/final")
     tokenizer.save_pretrained(f"{OUTPUT_DIR}/final")
-    print(f"Training complete. Model saved to {OUTPUT_DIR}/final")
+    print(f"✅ Saved to {OUTPUT_DIR}/final")
+
+    generate_plots(trainer)
+
+    print("\n" + "="*50)
+    print("NEXT: python training/evaluate_agent.py")
+    print("THEN: git add loss_curve.png reward_curve.png")
+    print("="*50)
 
 
 if __name__ == "__main__":
-    train()
+    train()