Update training/train_agent.py

training/train_agent.py

@@ -1,92 +1,98 @@
 """
 training/train_agent.py — SQL Database Engineer Agent
-Unsloth + GRPO training script.
-Run on venue GPU (April 25-26) with compute credits.
+Fixed version: reward_fn runs FULL EPISODES via /reset + /step
+This gives real delta rewards, milestones, and meaningful learning signal.
 
-FREE T4 (Colab):  MODEL_NAME=unsloth/Qwen2.5-1.5B-Instruct  (default)
-VENUE A100:       set ENV_VAR MODEL_NAME=unsloth/Qwen2.5-7B-Instruct
+FREE T4 (Colab/Kaggle): MODEL_NAME=unsloth/Qwen2.5-1.5B-Instruct
+VENUE A100:             MODEL_NAME=unsloth/Qwen2.5-7B-Instruct
 """
 
-import os
-import json
-import requests
-import sys
+import os, json, requests, sys, time
 from pathlib import Path
 
-# ── Try importing Unsloth (GPU only) ─────────────────────────
+# ── GPU check + imports ───────────────────────────────────────
+UNSLOTH_AVAILABLE = False
 try:
+    import torch
+    if not torch.cuda.is_available():
+        print("❌ No GPU found. Unsloth requires GPU.")
+        sys.exit(1)
     from unsloth import FastLanguageModel
     from trl import GRPOTrainer, GRPOConfig
-    import torch
+    from datasets import Dataset
     UNSLOTH_AVAILABLE = True
-    print("Unsloth + TRL loaded successfully")
-except ImportError:
-    UNSLOTH_AVAILABLE = False
-    print("Unsloth not available. Run: pip install unsloth trl")
-
-# ─────────────────────────────────────────────
-#  CONFIG — change MODEL_NAME via env var at venue
-# ─────────────────────────────────────────────
-
+    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"❌ Import error: {e}")
+    print("Run: pip install unsloth trl transformers datasets accelerate")
+    sys.exit(1)
+
+# ── 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/Qwen2.5-1.5B-Instruct")  # 1.5B for free T4
+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"))  # increase to 300+ at venue
-
-print(f"[CONFIG] Model:      {MODEL_NAME}")
-print(f"[CONFIG] Output:     {OUTPUT_DIR}")
-print(f"[CONFIG] Max steps:  {MAX_STEPS}")
-print(f"[CONFIG] ENV URL:    {ENV_URL}")
+MAX_STEPS  = int(os.getenv("MAX_STEPS", "100"))
 
-# ─────────────────────────────────────────────
-#  SYSTEM PROMPT
-# ─────────────────────────────────────────────
+print(f"\n[CONFIG] Model:     {MODEL_NAME}")
+print(f"[CONFIG] ENV URL:   {ENV_URL}")
+print(f"[CONFIG] Max steps: {MAX_STEPS}")
+print(f"[CONFIG] Output:    {OUTPUT_DIR}\n")
 
+# ── System prompt ─────────────────────────────────────────────
 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 have not inspected queries yet -> use inspect_query
-2. If you have not 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
-
-Respond with JSON only — no explanation, no markdown:
-{"action_type": "...", "payload": {...}}"""
-
-
-# ─────────────────────────────────────────────
-#  REWARD FUNCTION (calls live HF Space)
-# ─────────────────────────────────────────────
-
+Given the current database state, choose the BEST next action.
+
+Rules:
+1. First action MUST be inspect_query to see what's slow
+2. Then analyze_indexes to see what's missing
+3. Then create_index with correct table and columns
+4. Then analyze_statistics to update planner
+5. Finally submit_report when performance target is reached
+
+Respond with ONLY valid JSON — no markdown, no explanation:
+{"action_type": "create_index", "payload": {"table": "users", "columns": ["email"]}}"""
+
+# ── All 15 Round 2 scenario IDs ──────────────���────────────────
+ALL_SCENARIOS = [
+    "easy_s001", "easy_s002", "easy_s003", "easy_s004", "easy_s005",
+    "medium_s001", "medium_s002", "medium_s003", "medium_s004", "medium_s005",
+    "hard_s001", "hard_s002", "hard_s003", "hard_s004", "hard_s005",
+]
+
+# ── Parse LLM output → action dict ───────────────────────────
 def parse_action(text: str) -> dict:
-    """Parse LLM output into action dict. Returns safe fallback on failure."""
     try:
         text = text.strip()
-        if "```" in text:
-            text = text.split("```")[1]
-            if text.startswith("json"):
-                text = text[4:]
+        # Strip markdown
+        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:
+        if "action_type" in data and "payload" in data:
             return data
     except Exception:
         pass
-    # Safe fallback — never crashes
+    # Safe fallback
     return {"action_type": "inspect_query", "payload": {"query_id": "q1"}}
 
 
+# ── REWARD FUNCTION — runs FULL EPISODE ───────────────────────
 def reward_fn(prompts, completions, **kwargs):
     """
-    GRPO reward function — calls /grader on live environment.
-    Returns list of float rewards, one per completion.
-    Score always between 0.001 and 0.999.
+    KEY FIX: Runs a full episode per completion.
+    1. /reset with scenario
+    2. Parse LLM output as action
+    3. /step with that action
+    4. Get REAL reward including delta + milestones
+    5. /step with submit_report to get terminal score
+    Returns real rewards — not constant 0.5
     """
     rewards = []
-    task_ids = kwargs.get("task_ids", ["easy_s001"] * len(prompts))
+    batch = kwargs.get("batch", [])
 
     for i, (prompt, completion) in enumerate(zip(prompts, completions)):
         try:
@@ -96,28 +102,62 @@ def reward_fn(prompts, completions, **kwargs):
             else:
                 text = str(completion)
 
-            # Parse into action
+            # Pick scenario — rotate through all 15
+            scenario_id = ALL_SCENARIOS[i % len(ALL_SCENARIOS)]
+
+            # Parse LLM output
             action = parse_action(text)
-            task_id = task_ids[i] if i < len(task_ids) else "easy_s001"
-
-            # Call grader endpoint
-            resp = requests.post(
-                f"{ENV_URL}/grader",
-                json={"task_id": task_id, "action": action},
-                timeout=15,
-                headers={"Content-Type": "application/json"}
-            )
-            if resp.status_code == 200:
-                score = resp.json().get("score", 0.001)
-                score = max(0.001, min(0.999, float(score)))
+
+            # Step 1: Reset environment for this scenario
+            r = requests.post(f"{ENV_URL}/reset",
+                json={"task_id": scenario_id}, timeout=15)
+            if r.status_code != 200:
+                rewards.append(0.001)
+                continue
+
+            obs      = r.json()
+            baseline = obs.get("current_context", {}).get("performance_score", 0)
+
+            # Step 2: Submit the LLM's action
+            r2 = requests.post(f"{ENV_URL}/step",
+                json=action, timeout=15)
+            if r2.status_code != 200:
+                rewards.append(0.001)
+                continue
+
+            data       = r2.json()
+            step_score = data.get("reward", {}).get("score", 0.001)
+            db_delta   = data.get("info", {}).get("db_delta", 0)
+            perf_score = data.get("info", {}).get("performance_score", baseline)
+            milestones = data.get("info", {}).get("milestones", [])
+            done       = data.get("done", False)
+
+            # Step 3: If not done, submit report to get terminal score
+            if not done:
+                r3 = requests.post(f"{ENV_URL}/step",
+                    json={"action_type": "submit_report",
+                          "payload": {"summary": "Training episode complete."}},
+                    timeout=15)
+                if r3.status_code == 200:
+                    final_data = r3.json()
+                    terminal   = final_data.get("reward", {}).get("score", step_score)
+                    # Combine step reward + terminal
+                    final_score = (step_score * 0.4) + (terminal * 0.6)
+                else:
+                    final_score = step_score
             else:
-                score = 0.001
+                final_score = step_score
 
-            rewards.append(score)
-            print(f"  [REWARD] task={task_id} | action={action.get('action_type')} | score={score:.3f}")
+            # Clamp
+            final_score = max(0.001, min(0.999, final_score))
+            rewards.append(final_score)
+
+            print(f"  [REWARD] scenario={scenario_id} | "
+                  f"action={action.get('action_type')} | "
+                  f"db_delta=+{db_delta:.1f} | "
+                  f"milestones={milestones} | "
+                  f"score={final_score:.3f}")
 
-        except json.JSONDecodeError:
-            rewards.append(0.001)
         except Exception as e:
             print(f"  [REWARD] Error: {e}")
             rewards.append(0.001)
@@ -125,198 +165,149 @@ def reward_fn(prompts, completions, **kwargs):
     return rewards
 
 
-# ─────────────────────────────────────────────
-#  BUILD TRAINING DATASET
-# ─────────────────────────────────────────────
-
+# ── Build dataset with all 15 scenarios ───────────────────────
 def build_dataset():
-    """Build training examples from all 15 Round 2 scenarios."""
     scenarios = []
-
-    for fname in [
-        "dataset/easy_scenarios.json",
-        "dataset/medium_scenarios.json",
-        "dataset/hard_scenarios.json"
-    ]:
+    for fname in ["dataset/easy_scenarios.json",
+                  "dataset/medium_scenarios.json",
+                  "dataset/hard_scenarios.json"]:
         try:
             with open(fname) as f:
                 data = json.load(f)
                 scenarios.extend(data)
-                print(f"  Loaded {len(data)} scenarios from {fname}")
+                print(f"  Loaded {len(data)} from {fname}")
         except FileNotFoundError:
-            print(f"{fname} not found, skipping")
+            print(f"  ⚠️ {fname} not found")
 
     if not scenarios:
-        print("No local scenarios found. Fetching from live environment...")
-        try:
-            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]
-            print(f"  Fetched {len(scenarios)} tasks from HF Space")
-        except Exception as e:
-            print(f"Could not fetch tasks: {e}")
-            sys.exit(1)
+        print("  Fetching from live environment...")
+        resp = requests.get(f"{ENV_URL}/tasks", timeout=15)
+        tasks = resp.json().get("tasks", [])
+        scenarios = [t for t in tasks if t["id"].startswith(("easy_s","medium_s","hard_s"))]
 
     examples = []
-    for s in scenarios:
-        prompt = f"""{SYSTEM_PROMPT}
-
-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)}
-
-What is your next action?"""
-
+    for i, s in enumerate(scenarios):
+        # Build rich prompt with full DB state
+        tables_str = json.dumps(s.get("tables", []))
+        queries_str = json.dumps(s.get("slow_queries", []))
+        prompt = (
+            f"{SYSTEM_PROMPT}\n\n"
+            f"Scenario: {s.get('id')}\n"
+            f"Description: {s.get('description','')}\n"
+            f"Tables: {tables_str}\n"
+            f"Slow Queries: {queries_str}\n"
+            f"Performance: {s.get('performance_score_baseline',0)}/100 "
+            f"(target: {s.get('target_score',85)})\n\n"
+            f"What is your FIRST action?"
+        )
         examples.append({
-            "prompt":  prompt,
-            "task_id": s.get("id", "easy_s001"),
+            "prompt":      prompt,
+            "task_id":     s.get("id", ALL_SCENARIOS[i % len(ALL_SCENARIOS)]),
+            "scenario_id": s.get("id", ALL_SCENARIOS[i % len(ALL_SCENARIOS)]),
         })
 
-    print(f"  Built {len(examples)} training examples total")
-    from datasets import Dataset
+    print(f"  ✅ Dataset: {len(examples)} examples")
     return Dataset.from_list(examples)
 
 
-# ─────────────────────────────────────────────
-#  INFERENCE TEST — run immediately after save
-# ─────────────────────────────────────────────
-
-def test_inference(model, tokenizer):
-    """
-    REQUIRED: Test inference immediately after saving.
-    If this fails, the model was not saved correctly.
-    """
-    print("\n[INFERENCE TEST] Testing saved model...")
-    try:
-        FastLanguageModel.for_inference(model)
-
-        test_prompt = f"""{SYSTEM_PROMPT}
-
-Current Database State:
-- Scenario: easy_s001
-- Description: User lookup query taking 2s on 10K users table
-- Tables: [{{"name": "users", "rows": 10000, "indexes": ["PRIMARY"]}}]
-- Slow Queries: [{{"id": "q1", "sql": "SELECT * FROM users WHERE email=?", "avg_ms": 2000}}]
-- Performance Score: 8.0 / 100
-- Target Score: 80.0
-
-What is your next action?"""
-
-        inputs = tokenizer(
-            test_prompt,
-            return_tensors="pt",
-            truncation=True,
-            max_length=1024
-        ).to(model.device)
-
-        with torch.no_grad():
-            outputs = model.generate(
-                **inputs,
-                max_new_tokens  = 100,
-                temperature     = 0.3,
-                do_sample       = True,
-                pad_token_id    = tokenizer.eos_token_id,
-            )
-
-        response = tokenizer.decode(
-            outputs[0][inputs["input_ids"].shape[1]:],
-            skip_special_tokens=True
-        ).strip()
-
-        print(f"[INFERENCE TEST] Model output:\n  {response}")
-
-        # Validate output
-        action = parse_action(response)
-        print(f"[INFERENCE TEST] Parsed action: {action}")
-        print("[INFERENCE TEST] PASSED — model saved correctly!")
-        return True
-
-    except Exception as e:
-        print(f"[INFERENCE TEST] FAILED: {e}")
-        print("[INFERENCE TEST] Check model save path. Do NOT proceed without fixing this.")
-        return False
-
-
-# ─────────────────────────────────────────────
-#  MAIN TRAINING
-# ─────────────────────────────────────────────
+# ── Generate loss + reward curve from training logs ───────────
+def generate_training_plots(trainer):
+    import matplotlib
+    matplotlib.use("Agg")
+    import matplotlib.pyplot as plt
 
-def train():
-    if not UNSLOTH_AVAILABLE:
-        print(" Cannot train — Unsloth not installed or no GPU found")
-        print("Run: pip install unsloth trl transformers datasets accelerate")
+    logs    = [l for l in trainer.state.log_history if "loss" in l]
+    if not logs:
+        print("⚠️ No training logs found for plotting")
         return
 
-    print(f"\n Loading model: {MODEL_NAME}")
-    print(f" Environment:   {ENV_URL}\n")
-
-    # Verify environment is reachable
+    steps   = [l.get("step", i)   for i, l in enumerate(logs)]
+    losses  = [l.get("loss", 0)   for l in logs]
+    rewards = [l.get("reward", l.get("train_loss", 0)) for l in logs]
+
+    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 4))
+    fig.suptitle("GRPO Training — SQL Database Engineer Agent",
+                 fontsize=13, fontweight="bold")
+
+    ax1.plot(steps, losses, "b-o", lw=2, ms=4, label="Loss")
+    ax1.set_xlabel("Training Step")
+    ax1.set_ylabel("Loss")
+    ax1.set_title("Training Loss (↓ = learning)")
+    ax1.grid(True, alpha=0.3)
+    ax1.legend()
+
+    ax2.plot(steps, rewards, "g-o", lw=2, ms=4, label="Reward")
+    ax2.set_xlabel("Training Step")
+    ax2.set_ylabel("Reward")
+    ax2.set_title("Reward During Training (↑ = improving)")
+    ax2.grid(True, alpha=0.3)
+    ax2.legend()
+
+    plt.tight_layout()
+    plt.savefig("loss_curve.png", dpi=150, bbox_inches="tight")
+    print("✅ loss_curve.png saved")
+
+    # Print summary
+    if losses:
+        print(f"  Loss:   {losses[0]:.4f} → {losses[-1]:.4f}")
+    if rewards:
+        valid_r = [r for r in rewards if r > 0]
+        if valid_r:
+            print(f"  Reward: {valid_r[0]:.4f} → {valid_r[-1]:.4f}")
+
+
+# ── Main training ─────────────────────────────────────────────
+def train():
+    # Verify environment
     try:
         r = requests.get(f"{ENV_URL}/health", timeout=10)
-        version = r.json().get("version", "?")
-        print(f" Environment reachable — version {version}")
+        print(f"✅ Environment live: v{r.json().get('version','?')}\n")
     except Exception as e:
-        print(f" Cannot reach environment at {ENV_URL}: {e}")
-        print("Check ENV_URL and make sure HF Space is running.")
+        print(f"❌ Cannot reach {ENV_URL}: {e}")
         sys.exit(1)
 
-    # ── Load model ───────────────────────────────────────────
+    # Load model
+    print(f"⏳ Loading {MODEL_NAME}...")
     model, tokenizer = FastLanguageModel.from_pretrained(
         model_name     = MODEL_NAME,
         max_seq_length = 2048,
-        load_in_4bit   = True,   # QLoRA — required for T4
-        dtype          = None,   # Auto detect
+        load_in_4bit   = True,
         token          = HF_TOKEN or None,
     )
-    print(" Model loaded")
-
-    # ── Apply 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"],
+        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(" LoRA adapters applied")
+    print("✅ Model + LoRA ready\n")
 
-    # ── Build dataset ────────────────────────────────────────
-    print("\n[DATASET] Building training dataset...")
+    # Dataset
+    print("⏳ Building dataset...")
     dataset = build_dataset()
-    print(f" Dataset ready: {len(dataset)} examples")
 
-    # ── Reward wrapper ───────────────────────────────────────
+    # Reward wrapper
     def reward_wrapper(prompts, completions, **kwargs):
-        batch = kwargs.get("batch", [])
-        if batch and hasattr(batch[0], "get"):
-            task_ids = [b.get("task_id", "easy_s001") for b in batch]
-        else:
-            task_ids = ["easy_s001"] * len(prompts)
-        return reward_fn(prompts, completions, task_ids=task_ids)
-
-    # ── GRPO config ──────────────────────────────────────────
-    # NOTE: batch_size=1, num_generations=2 for free T4
-    # At venue A100: increase to batch_size=2, num_generations=4
+        return reward_fn(prompts, completions, **kwargs)
+
+    # GRPO config
     config = GRPOConfig(
         output_dir                  = OUTPUT_DIR,
         max_steps                   = MAX_STEPS,
-        per_device_train_batch_size = 1,   # 1 for T4, 2 for A100
-        gradient_accumulation_steps = 8,
+        per_device_train_batch_size = 1,
+        gradient_accumulation_steps = 4,
         learning_rate               = 5e-6,
-        max_completion_length       = 256,
-        num_generations             = 2,   # 2 for T4, 4 for A100
-        temperature                 = 0.8,
-        logging_steps               = 5,
-        save_steps                  = 50,
-        save_total_limit            = 2,
+        max_completion_length       = 200,
+        num_generations             = 2,
+        temperature                 = 0.9,
+        logging_steps               = 1,
+        save_steps                  = 25,
+        save_total_limit            = 3,
         warmup_ratio                = 0.1,
         report_to                   = "none",
         remove_unused_columns       = False,
@@ -330,50 +321,27 @@ def train():
         train_dataset = dataset,
     )
 
-    # ── Train ────────────────────────────────────────────────
-    print(f"\n🏋️  Starting GRPO training — {MAX_STEPS} steps...")
-    print("Watch the 'reward' column — it should increase over time.\n")
+    print(f"🏋️  Starting GRPO — {MAX_STEPS} steps")
+    print("Watch for: db_delta > 0 and milestones in logs\n")
     trainer.train()
-    print("\n Training complete!")
+    print("\n✅ Training complete!")
 
-    # ── Save — ADAPTER ONLY (correct way for QLoRA) ──────────
-    # DO NOT call merge_and_unload() on 4-bit model
-    # DO NOT upcast to 16-bit and merge naively
-    # CORRECT: save adapter weights only, load with from_pretrained later
-    print(f"\n[SAVE] Saving adapter to {OUTPUT_DIR}/final ...")
+    # Save
     Path(OUTPUT_DIR).mkdir(parents=True, exist_ok=True)
     model.save_pretrained(f"{OUTPUT_DIR}/final")
     tokenizer.save_pretrained(f"{OUTPUT_DIR}/final")
-
-    # Save config for reference
-    with open(f"{OUTPUT_DIR}/final/training_config.json", "w") as f:
-        json.dump({
-            "model_name":   MODEL_NAME,
-            "max_steps":    MAX_STEPS,
-            "save_method":  "adapter_only_qlora",
-            "lora_r":       16,
-            "lora_alpha":   16,
-        }, f, indent=2)
-    print(f" Adapter saved to {OUTPUT_DIR}/final")
-
-    # ── IMMEDIATE inference test (required) ──────────────────
-    passed = test_inference(model, tokenizer)
-
-    # ── Summary ──────────────────────────────────────────────
-    print("\n" + "="*60)
-    print("TRAINING COMPLETE")
-    print("="*60)
-    print(f"  Model:         {MODEL_NAME}")
-    print(f"  Steps:         {MAX_STEPS}")
-    print(f"  Saved to:      {OUTPUT_DIR}/final")
-    print(f"  Save method:   Adapter only (QLoRA safe)")
-    print(f"  Inference test: {' PASSED' if passed else ' FAILED'}")
-    print("="*60)
-    print("\nNext steps:")
-    print("  1. python training/evaluate_agent.py")
-    print("  2. Open reward_curve.png — show to judges")
-    print("  3. git add reward_curve.png && git commit && git push")
-    print("="*60)
+    print(f"✅ Saved to {OUTPUT_DIR}/final")
+
+    # Generate plots
+    generate_training_plots(trainer)
+
+    print("\n" + "="*55)
+    print("NEXT STEPS:")
+    print("  python training/evaluate_agent.py")
+    print("  git add loss_curve.png reward_curve.png")
+    print("  git commit -m 'Real training evidence'")
+    print("  git push origin main")
+    print("="*55)
 
 
 if __name__ == "__main__":