synth-ai 0.2.4.dev8__py3-none-any.whl → 0.2.5__py3-none-any.whl

synth_ai/environments/examples/crafter_classic/agent_demos/crafter_modal_ft/kick_off_ft_modal.py

@@ -0,0 +1,384 @@
+#!/usr/bin/env python3
+"""
+Modal/Synth Fine-Tuning Script
+==============================
+Uploads a JSONL file to Modal, kicks off a fine-tuning job, and polls until completion.
+Updated for OpenAI v1 compatible unified fine-tuning service.
+"""
+
+import argparse
+import asyncio
+import json
+import os
+import random
+import sys
+import time
+from datetime import datetime
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import httpx
+
+# Add synth_ai to path (optional - only if needed)
+# sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent.parent.parent))
+
+# Import Synth LM utilities (optional - only if needed)
+# from synth_ai.lm import SynthConfig
+
+# Modal fine-tuning endpoints - Updated for OpenAI v1 compatible service
+MODAL_BASE_URL = os.getenv('MODAL_BASE_URL', os.getenv('SYNTH_BASE_URL', 'https://synth-laboratories--unified-ft-service-fastapi-app.modal.run'))
+MODAL_API_KEY = os.getenv('MODAL_API_KEY', os.getenv('SYNTH_API_KEY', 'sk-test-11111111111111111111111111111111'))
+
+
+def analyze_jsonl_tokens(file_path: Path, model: str) -> tuple[int, int, float]:
+    """Analyze JSONL file to estimate token usage."""
+    print(f"🔍 Analyzing {file_path.name} for token usage...")
+    
+    # For Modal/Synth, we'll do a rough estimate based on character count
+    # Approximate: 1 token ≈ 4 characters (rough estimate)
+    CHARS_PER_TOKEN = 4  # noqa: N806
+    
+    total_input_tokens = 0
+    total_output_tokens = 0
+    line_count = 0
+    
+    with open(file_path, 'r') as f:
+        for line in f:
+            try:
+                data = json.loads(line.strip())
+                messages = data.get('messages', [])
+                
+                # Count input tokens (all messages except the last assistant message)
+                input_chars = 0
+                output_chars = 0
+                
+                for i, msg in enumerate(messages):
+                    content = msg.get('content', '')
+                    char_count = len(content)
+                    
+                    if msg.get('role') == 'assistant' and i == len(messages) - 1:
+                        # This is the target output
+                        output_chars += char_count
+                    else:
+                        # This is input context
+                        input_chars += char_count
+                    
+                    # Include tool calls
+                    tool_calls = msg.get('tool_calls', [])
+                    for tc in tool_calls:
+                        if tc.get('function', {}).get('arguments'):
+                            if i == len(messages) - 1 and msg.get('role') == 'assistant':
+                                output_chars += len(tc['function']['arguments'])
+                            else:
+                                input_chars += len(tc['function']['arguments'])
+                
+                input_tokens = input_chars // CHARS_PER_TOKEN
+                output_tokens = output_chars // CHARS_PER_TOKEN
+                
+                total_input_tokens += input_tokens
+                total_output_tokens += output_tokens
+                line_count += 1
+                
+            except json.JSONDecodeError:
+                print(f"   ⚠️  Skipping invalid JSON line {line_count + 1}")
+                continue
+            except Exception as e:
+                print(f"   ⚠️  Error processing line {line_count + 1}: {e}")
+                continue
+    
+    total_tokens = total_input_tokens + total_output_tokens
+    avg_tokens_per_line = total_tokens / line_count if line_count > 0 else 0
+    
+    print(f"   📊 Analysis complete:")
+    print(f"      Lines: {line_count:,}")
+    print(f"      Input tokens (est.): {total_input_tokens:,}")
+    print(f"      Output tokens (est.): {total_output_tokens:,}")
+    print(f"      Total tokens (est.): {total_tokens:,}")
+    print(f"      Avg tokens/line: {avg_tokens_per_line:.1f}")
+    
+    return line_count, total_tokens, avg_tokens_per_line
+
+
+def create_subset_file(original_path: Path, num_lines: int) -> Path:
+    """Create a subset of the original JSONL file with specified number of lines."""
+    subset_path = original_path.parent / f"{original_path.stem}_subset_{num_lines}.jsonl"
+    
+    print(f"📝 Creating subset with {num_lines} lines...")
+    
+    # Read all lines
+    with open(original_path, 'r') as f:
+        all_lines = [line.strip() for line in f if line.strip()]
+    
+    # Randomly sample lines
+    if num_lines >= len(all_lines):
+        selected_lines = all_lines
+        print(f"   ⚠️  Requested {num_lines} lines, but file only has {len(all_lines)}. Using all lines.")
+    else:
+        selected_lines = random.sample(all_lines, num_lines)
+    
+    # Write subset
+    with open(subset_path, 'w') as f:
+        for line in selected_lines:
+            f.write(line + '\n')
+    
+    print(f"   ✅ Subset saved to: {subset_path.name}")
+    return subset_path
+
+
+async def upload_file(file_path: Path) -> str:
+    """Upload training file to Modal using OpenAI v1 compatible endpoint."""
+    print(f"📤 Uploading {file_path.name} ({file_path.stat().st_size / 1024 / 1024:.1f} MB)...")
+    
+    async with httpx.AsyncClient(timeout=300.0) as client:
+        # Read file content
+        with open(file_path, 'rb') as f:
+            file_content = f.read()
+        
+        # Create multipart form data for OpenAI v1 compatible endpoint
+        files = {
+            'file': (file_path.name, file_content, 'application/json')
+        }
+        
+        # Use OpenAI v1 compatible endpoint
+        response = await client.post(
+            f"{MODAL_BASE_URL}/v1/files?purpose=fine-tune",
+            files=files,
+            headers={"Authorization": f"Bearer {MODAL_API_KEY}"}
+        )
+        
+        if response.status_code != 200:
+            raise Exception(f"Upload failed: {response.status_code} - {response.text}")
+        
+        result = response.json()
+        file_id = result.get('id')
+        
+        print(f"✅ File uploaded: {file_id}")
+        return file_id
+
+
+async def create_fine_tune_job(file_id: str, model: str = "Qwen/Qwen2.5-7B-Instruct", 
+                              config: Optional[Dict[str, Any]] = None) -> str:
+    """Create a fine-tuning job on Modal using OpenAI v1 compatible endpoint."""
+    print(f"🚀 Starting fine-tune job for {model}...")
+    
+    # Default fine-tuning configuration for OpenAI v1 compatible service
+    ft_config = {
+        "model": model,
+        "training_file": file_id,
+        "training_type": "sft",  # sft or dpo
+        "hyperparameters": {
+            "n_epochs": 3,
+            "batch_size": 4,
+            "learning_rate": 5e-5,
+            "use_qlora": True,
+            "lora_r": 16,
+            "lora_alpha": 32,
+            "lora_dropout": 0.1,
+        },
+        "suffix": f"modal-{int(time.time())}"
+    }
+    
+    # Update with user config if provided
+    if config:
+        ft_config["hyperparameters"].update(config)
+    
+    async with httpx.AsyncClient(timeout=300.0) as client:
+        response = await client.post(
+            f"{MODAL_BASE_URL}/v1/fine_tuning/jobs",
+            json=ft_config,
+            headers={
+                "Authorization": f"Bearer {MODAL_API_KEY}",
+                "Content-Type": "application/json"
+            }
+        )
+        
+        if response.status_code != 200:
+            raise Exception(f"Job creation failed: {response.status_code} - {response.text}")
+        
+        result = response.json()
+        job_id = result.get('id')
+        
+        print(f"✅ Fine-tune job created: {job_id}")
+        print(f"   Model: {model}")
+        print(f"   Status: {result.get('status', 'created')}")
+        
+        return job_id
+
+
+async def poll_job_status(job_id: str, poll_interval: int = 30) -> Optional[str]:
+    """Poll job status until completion using OpenAI v1 compatible endpoint."""
+    print(f"⏳ Polling job {job_id} every {poll_interval}s...")
+    
+    start_time = time.time()
+    last_status = None
+    
+    async with httpx.AsyncClient(timeout=60.0) as client:
+        while True:
+            try:
+                response = await client.get(
+                    f"{MODAL_BASE_URL}/v1/fine_tuning/jobs/{job_id}",
+                    headers={"Authorization": f"Bearer {MODAL_API_KEY}"}
+                )
+                
+                if response.status_code != 200:
+                    print(f"   ⚠️  Failed to get status: {response.status_code}")
+                    await asyncio.sleep(poll_interval)
+                    continue
+                
+                job = response.json()
+                status = job.get('status', 'unknown')
+                
+                if status != last_status:
+                    elapsed = time.time() - start_time
+                    print(f"   Status: {status} (elapsed: {elapsed/60:.1f}m)")
+                    last_status = status
+                    
+                    # Show training progress if available
+                    if 'hyperparameters' in job:
+                        hp = job['hyperparameters']
+                        print(f"   Model: {job.get('model', 'unknown')}")
+                        print(f"   Training file: {job.get('training_file', 'unknown')}")
+                
+                # Terminal states
+                if status == "succeeded":
+                    print(f"🎉 Fine-tuning completed successfully!")
+                    final_model = job.get('fine_tuned_model')
+                    if final_model:
+                        print(f"   Final model: {final_model}")
+                    return final_model
+                
+                elif status == "failed":
+                    print(f"❌ Fine-tuning failed!")
+                    if 'error' in job:
+                        print(f"   Error: {job['error']}")
+                    return None
+                
+                elif status == "cancelled":
+                    print(f"⚠️  Fine-tuning was cancelled")
+                    return None
+                
+                # Continue polling for running states
+                elif status in ["queued", "running", "validating_files"]:
+                    await asyncio.sleep(poll_interval)
+                    continue
+                
+                else:
+                    print(f"⚠️  Unknown status: {status}")
+                    await asyncio.sleep(poll_interval)
+                    continue
+                    
+            except KeyboardInterrupt:
+                print(f"\n⚠️  Interrupted by user. Job {job_id} is still running on Modal.")
+                print(f"   Check status later with the job ID: {job_id}")
+                return None
+            
+            except Exception as e:
+                print(f"❌ Error polling job: {e}")
+                await asyncio.sleep(poll_interval)
+                continue
+
+
+async def main():
+    parser = argparse.ArgumentParser(description="Modal/Synth Fine-Tuning Script")
+    parser.add_argument("jsonl_file", type=Path, help="Path to JSONL training file")
+    parser.add_argument("--model", default="Qwen/Qwen2.5-7B-Instruct", 
+                       help="Base model to fine-tune (default: Qwen/Qwen2.5-7B-Instruct)")
+    parser.add_argument("--epochs", type=int, default=3, help="Number of training epochs")
+    parser.add_argument("--batch-size", type=int, default=4, help="Training batch size")
+    parser.add_argument("--learning-rate", type=float, default=5e-5, help="Learning rate")
+    parser.add_argument("--poll-interval", type=int, default=30, 
+                       help="Polling interval in seconds (default: 30)")
+    parser.add_argument("--subset", type=int, help="Use a random subset of N lines")
+    parser.add_argument("--training-type", default="sft", choices=["sft", "dpo"],
+                       help="Training type: sft (supervised) or dpo (preference)")
+    
+    args = parser.parse_args()
+    
+    # Validate file
+    if not args.jsonl_file.exists():
+        print(f"❌ File not found: {args.jsonl_file}")
+        sys.exit(1)
+    
+    if not args.jsonl_file.suffix == '.jsonl':
+        print(f"⚠️  Warning: File doesn't have .jsonl extension: {args.jsonl_file}")
+    
+    # Check API key
+    if not MODAL_API_KEY:
+        print("❌ Modal API key required. Set MODAL_API_KEY or SYNTH_API_KEY env var")
+        sys.exit(1)
+    
+    # Analyze tokens first
+    line_count, total_tokens, avg_tokens = analyze_jsonl_tokens(args.jsonl_file, args.model)
+    
+    # Estimate cost (Modal/Synth pricing varies by model)
+    # This is a rough estimate - actual costs depend on Modal pricing
+    estimated_hours = total_tokens / 1_000_000 * 0.5  # Rough estimate
+    print(f"\n⏱️  Estimated training time: ~{estimated_hours:.1f} hours")
+    print(f"   (Based on ~2M tokens/hour - actual time may vary)")
+    
+    # Use subset if requested
+    training_file = args.jsonl_file
+    if args.subset:
+        if args.subset > line_count:
+            print(f"⚠️  Subset size ({args.subset}) exceeds file size ({line_count}). Using all lines.")
+        else:
+            training_file = create_subset_file(args.jsonl_file, args.subset)
+    
+    print("\n🤖 Modal/Synth Fine-Tuning Pipeline")
+    print("=" * 50)
+    print(f"Training file: {training_file}")
+    print(f"Base model: {args.model}")
+    print(f"Training type: {args.training_type}")
+    print(f"Epochs: {args.epochs}")
+    print(f"Batch size: {args.batch_size}")
+    print(f"Learning rate: {args.learning_rate}")
+    print("=" * 50)
+    
+    try:
+        # Step 1: Upload file
+        file_id = await upload_file(training_file)
+        
+        # Step 2: Create fine-tune job
+        ft_config = {
+            "n_epochs": args.epochs,
+            "batch_size": args.batch_size,
+            "learning_rate": args.learning_rate,
+        }
+        job_id = await create_fine_tune_job(file_id, args.model, ft_config)
+        
+        # Step 3: Poll until completion
+        final_model = await poll_job_status(job_id, args.poll_interval)
+        
+        if final_model:
+            print("\n" + "=" * 50)
+            print(f"🎯 SUCCESS! Fine-tuned model ready: {final_model}")
+            print("=" * 50)
+            
+            # Show usage example
+            print("\n📝 Usage example:")
+            print(f'import httpx')
+            print(f'')
+            print(f'async def test_model():')
+            print(f'    async with httpx.AsyncClient() as client:')
+            print(f'        response = await client.post(')
+            print(f'            "{MODAL_BASE_URL}/v1/chat/completions",')
+            print(f'            headers={{"Authorization": f"Bearer {MODAL_API_KEY}"}},')
+            print(f'            json={{')
+            print(f'                "model": "{final_model}",')
+            print(f'                "messages": [{{"role": "user", "content": "Hello!"}}]')
+            print(f'            }}')
+            print(f'        )')
+            print(f'        return response.json()')
+        else:
+            print("\n❌ Fine-tuning did not complete successfully")
+            sys.exit(1)
+            
+    except Exception as e:
+        print(f"\n❌ Unexpected error: {e}")
+        import traceback
+        traceback.print_exc()
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

synth_ai/environments/examples/crafter_classic/agent_demos/crafter_modal_ft/old/analyze_action_results.py

@@ -0,0 +1,53 @@
+#!/usr/bin/env python3
+"""Analyze why actions aren't producing expected results."""
+
+import json
+import glob
+from pathlib import Path
+
+# Find the latest results file
+result_files = glob.glob("crafter_lm_synth_results_*.json")
+if not result_files:
+    print("No result files found")
+    exit(1)
+
+latest_file = max(result_files, key=lambda f: Path(f).stat().st_mtime)
+print(f"📊 Analyzing: {latest_file}\n")
+
+with open(latest_file) as f:
+    data = json.load(f)
+
+# Check each episode
+for ep_idx, episode in enumerate(data.get('results', [])):
+    if 'error' in episode:
+        continue
+        
+    print(f"\n{'='*60}")
+    print(f"EPISODE {ep_idx}")
+    print(f"{'='*60}")
+    
+    steps = episode.get('step_results', [])
+    
+    # Track inventory changes
+    prev_inventory = {}
+    
+    for step in steps:
+        turn = step.get('turn', 0)
+        action = step.get('action', 'unknown')
+        reward = step.get('reward', 0)
+        
+        # This won't work because we don't store obs in step_results
+        # We need to look at the actual traces or add obs to step_results
+        
+    achievements = episode.get('achievements_unlocked', [])
+    print(f"\nFinal achievements: {achievements}")
+    print(f"Total reward: {episode.get('total_reward', 0)}")
+    print(f"Steps: {episode.get('steps', 0)}")
+
+# Summary
+summary = data.get('summary', {})
+print(f"\n{'='*60}")
+print("OVERALL SUMMARY")
+print(f"{'='*60}")
+print(f"Unique achievements: {summary.get('unique_achievements', [])}")
+print(f"Average reward: {summary.get('avg_reward', 0)}")

synth_ai/environments/examples/crafter_classic/agent_demos/crafter_modal_ft/old/analyze_agent_actions.py

@@ -0,0 +1,178 @@
+#!/usr/bin/env python3
+"""Analyze agent actions and model responses from trace data."""
+
+import duckdb
+import json
+from pathlib import Path
+from collections import Counter
+
+def analyze_agent_actions(db_path: str):
+    """Analyze agent actions and responses."""
+    conn = duckdb.connect(db_path, read_only=True)
+    
+    print("🔍 Analyzing agent actions and responses...\n")
+    
+    # Get all events
+    all_events = conn.execute("""
+        SELECT 
+            e.session_id,
+            e.event_type,
+            e.event_time,
+            e.metadata,
+            e.system_state_after,
+            e.system_state_before,
+            e.reward,
+            e.terminated
+        FROM events e
+        ORDER BY e.session_id, e.event_time
+    """).fetchall()
+    
+    print(f"Total events found: {len(all_events)}\n")
+    
+    # Group events by session
+    sessions = {}
+    for event in all_events:
+        session_id = event[0]
+        if session_id not in sessions:
+            sessions[session_id] = []
+        sessions[session_id].append(event)
+    
+    print(f"Total sessions: {len(sessions)}\n")
+    
+    # Analyze each session
+    for i, (session_id, events) in enumerate(sessions.items()):
+        if i >= 3:  # Only analyze first 3 sessions
+            break
+            
+        print(f"\n{'='*60}")
+        print(f"SESSION {i+1}: {session_id}")
+        print(f"{'='*60}")
+        print(f"Total events in session: {len(events)}")
+        
+        # Track actions and achievements
+        actions_taken = []
+        achievements_unlocked = set()
+        total_reward = 0
+        
+        for event in events:
+            event_type = event[1]
+            metadata = json.loads(event[3]) if event[3] else {}
+            state_after = json.loads(event[4]) if event[4] else {}
+            state_before = json.loads(event[5]) if event[5] else {}
+            reward = event[6] or 0
+            terminated = event[7]
+            
+            total_reward += reward
+            
+            # Track runtime events (actions)
+            if event_type == 'runtime' and 'action_name' in metadata:
+                action = metadata['action_name']
+                actions_taken.append(action)
+                
+            # Check for achievements
+            if 'public_state' in state_after:
+                public_state = state_after['public_state']
+                if 'achievements_status' in public_state:
+                    for ach, unlocked in public_state['achievements_status'].items():
+                        if unlocked:
+                            achievements_unlocked.add(ach)
+            
+            # Look for generation responses
+            if event_type == 'generation' and metadata:
+                print(f"\n--- Generation Event ---")
+                if 'model' in metadata:
+                    print(f"Model: {metadata['model']}")
+                if 'response' in metadata:
+                    response = metadata['response']
+                    if isinstance(response, str):
+                        print(f"Response preview: {response[:200]}")
+                    elif isinstance(response, dict):
+                        if 'content' in response:
+                            print(f"Content: {response['content'][:200]}")
+                        if 'tool_calls' in response:
+                            print(f"Tool calls: {response['tool_calls']}")
+        
+        # Summary for this session
+        print(f"\n--- Session Summary ---")
+        print(f"Total reward: {total_reward}")
+        print(f"Achievements unlocked: {achievements_unlocked if achievements_unlocked else 'None'}")
+        print(f"Total actions taken: {len(actions_taken)}")
+        
+        if actions_taken:
+            action_counts = Counter(actions_taken)
+            print(f"\nAction distribution:")
+            for action, count in action_counts.most_common(10):
+                print(f"  {action}: {count}")
+    
+    # Overall analysis
+    print(f"\n\n{'='*60}")
+    print("OVERALL ANALYSIS")
+    print(f"{'='*60}")
+    
+    # Check for any achievements across all sessions
+    all_achievements = conn.execute("""
+        SELECT DISTINCT
+            json_extract_string(system_state_after, '$.public_state.achievements_status') as achievements
+        FROM events
+        WHERE system_state_after IS NOT NULL
+        AND json_extract_string(system_state_after, '$.public_state.achievements_status') IS NOT NULL
+        LIMIT 10
+    """).fetchall()
+    
+    print(f"\nChecking achievement states...")
+    total_unlocked = 0
+    for ach_json, in all_achievements[:3]:
+        if ach_json:
+            achievements = json.loads(ach_json)
+            unlocked = [k for k, v in achievements.items() if v]
+            if unlocked:
+                total_unlocked += len(unlocked)
+                print(f"Found unlocked: {unlocked}")
+    
+    if total_unlocked == 0:
+        print("❌ No achievements were unlocked in any session!")
+    
+    # Check for specific issues
+    print(f"\n\n{'='*60}")
+    print("POTENTIAL ISSUES")
+    print(f"{'='*60}")
+    
+    # Look for model responses
+    model_responses = conn.execute("""
+        SELECT COUNT(*) 
+        FROM events 
+        WHERE event_type = 'generation'
+    """).fetchone()[0]
+    
+    print(f"Generation events found: {model_responses}")
+    
+    if model_responses == 0:
+        print("❌ No generation events found - agent may not be responding")
+    
+    # Look for specific metadata patterns
+    print("\nChecking event metadata patterns...")
+    metadata_samples = conn.execute("""
+        SELECT event_type, metadata
+        FROM events
+        WHERE metadata IS NOT NULL
+        AND metadata != '{}'
+        AND metadata != '[]'
+        LIMIT 20
+    """).fetchall()
+    
+    for event_type, metadata_str in metadata_samples[:5]:
+        metadata = json.loads(metadata_str)
+        print(f"\n{event_type}: {list(metadata.keys())}")
+        if 'action_name' in metadata:
+            print(f"  Action: {metadata['action_name']}")
+        if 'model' in metadata:
+            print(f"  Model: {metadata['model']}")
+    
+    conn.close()
+
+if __name__ == "__main__":
+    db_path = "./traces_v2_synth/traces.duckdb"
+    if Path(db_path).exists():
+        analyze_agent_actions(db_path)
+    else:
+        print(f"❌ Database not found at {db_path}")