synth-ai 0.2.13.dev2__py3-none-any.whl → 0.2.16__py3-none-any.whl

examples/qwen_vl/test_image_validation.py

@@ -0,0 +1,201 @@
+#!/usr/bin/env python3
+"""
+Quick test script to demonstrate image validation.
+
+Run from synth-ai root:
+    uv run python examples/qwen_vl/test_image_validation.py
+"""
+
+from synth_ai.learning.sft.data import coerce_example, validate_vision_example
+
+# Test cases
+test_cases = [
+    {
+        "name": "Valid - HTTP URL",
+        "data": {
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": "Describe this"},
+                        {"type": "image_url", "image_url": {"url": "https://example.com/image.jpg"}},
+                    ],
+                },
+                {"role": "assistant", "content": "A beautiful image"},
+            ]
+        },
+        "should_pass": True,
+    },
+    {
+        "name": "Valid - Base64",
+        "data": {
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KGgo..."}},
+                    ],
+                },
+                {"role": "assistant", "content": "An image"},
+            ]
+        },
+        "should_pass": True,
+    },
+    {
+        "name": "Invalid - Empty URL",
+        "data": {
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": "What's this?"},
+                        {"type": "image_url", "image_url": {"url": ""}},  # Empty!
+                    ],
+                },
+                {"role": "assistant", "content": "Response"},
+            ]
+        },
+        "should_pass": False,
+    },
+    {
+        "name": "Invalid - Missing URL field",
+        "data": {
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "image_url", "image_url": {}},  # No url field!
+                    ],
+                },
+                {"role": "assistant", "content": "Response"},
+            ]
+        },
+        "should_pass": False,
+    },
+    {
+        "name": "Invalid - Null URL",
+        "data": {
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "image_url", "image_url": {"url": None}},  # Null!
+                    ],
+                },
+                {"role": "assistant", "content": "Response"},
+            ]
+        },
+        "should_pass": False,
+    },
+    {
+        "name": "Invalid - Whitespace URL",
+        "data": {
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "image_url", "image_url": {"url": "   "}},  # Whitespace!
+                    ],
+                },
+                {"role": "assistant", "content": "Response"},
+            ]
+        },
+        "should_pass": False,
+    },
+    {
+        "name": "Invalid - Mixed valid and invalid",
+        "data": {
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "image_url", "image_url": {"url": "https://example.com/valid.jpg"}},
+                        {"type": "image_url", "image_url": {"url": ""}},  # One invalid!
+                    ],
+                },
+                {"role": "assistant", "content": "Response"},
+            ]
+        },
+        "should_pass": False,
+    },
+    {
+        "name": "Invalid - Non-string URL",
+        "data": {
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "image_url", "image_url": {"url": 12345}},  # Integer!
+                    ],
+                },
+                {"role": "assistant", "content": "Response"},
+            ]
+        },
+        "should_pass": False,
+    },
+]
+
+
+def main():
+    print("=" * 80)
+    print("IMAGE VALIDATION TEST")
+    print("=" * 80)
+    print()
+    
+    passed = 0
+    failed = 0
+    
+    for test in test_cases:
+        name = test["name"]
+        data = test["data"]
+        should_pass = test["should_pass"]
+        
+        try:
+            example = coerce_example(data)
+            is_valid, error = validate_vision_example(example, require_images=True)
+            
+            if should_pass:
+                if is_valid:
+                    print(f"✅ PASS: {name}")
+                    print(f"   → Correctly accepted valid example")
+                    passed += 1
+                else:
+                    print(f"❌ FAIL: {name}")
+                    print(f"   → Should pass but got error: {error}")
+                    failed += 1
+            else:
+                if not is_valid:
+                    print(f"✅ PASS: {name}")
+                    print(f"   → Correctly rejected: {error}")
+                    passed += 1
+                else:
+                    print(f"❌ FAIL: {name}")
+                    print(f"   → Should fail but passed validation")
+                    failed += 1
+        except Exception as exc:
+            if should_pass:
+                print(f"❌ FAIL: {name}")
+                print(f"   → Unexpected exception: {exc}")
+                failed += 1
+            else:
+                print(f"✅ PASS: {name}")
+                print(f"   → Correctly raised exception: {exc}")
+                passed += 1
+        
+        print()
+    
+    print("=" * 80)
+    print(f"RESULTS: {passed}/{len(test_cases)} passed, {failed}/{len(test_cases)} failed")
+    print("=" * 80)
+    
+    if failed == 0:
+        print("🎉 All tests passed!")
+        return 0
+    else:
+        print(f"⚠️  {failed} test(s) failed")
+        return 1
+
+
+if __name__ == "__main__":
+    exit(main())
+

examples/qwen_vl/test_sft_vision_data.py

@@ -0,0 +1,110 @@
+"""Generate test vision SFT dataset for Qwen3-VL-2B."""
+
+import base64
+import json
+from pathlib import Path
+from io import BytesIO
+
+try:
+    from PIL import Image
+except ImportError:
+    print("❌ PIL not available")
+    exit(1)
+
+BASE_DIR = Path(__file__).resolve().parent
+
+def create_test_image(color: str) -> str:
+    """Create a 64x64 colored square and return base64 data URL."""
+    colors = {
+        "red": (255, 0, 0),
+        "blue": (0, 0, 255),
+        "green": (0, 255, 0),
+        "yellow": (255, 255, 0),
+        "purple": (128, 0, 128),
+    }
+    
+    img = Image.new('RGB', (64, 64), color=colors[color])
+    buffer = BytesIO()
+    img.save(buffer, format='PNG')
+    b64 = base64.b64encode(buffer.getvalue()).decode('utf-8')
+    return f"data:image/png;base64,{b64}"
+
+
+def main():
+    output_dir = BASE_DIR / "test_data"
+    output_dir.mkdir(parents=True, exist_ok=True)
+    
+    output_file = output_dir / "vision_sft_test.jsonl"
+    
+    # Create 10 training examples with different colored images
+    examples = []
+    colors = ["red", "blue", "green", "yellow", "purple"]
+    
+    for i, color in enumerate(colors):
+        # Simple color identification
+        examples.append({
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": "What color is this image? Answer in one word."},
+                        {"type": "image_url", "image_url": {"url": create_test_image(color)}},
+                    ],
+                },
+                {
+                    "role": "assistant",
+                    "content": color.capitalize(),
+                },
+            ],
+            "metadata": {"example_id": f"color_{i}", "type": "color_id"},
+        })
+        
+        # Describe the image
+        examples.append({
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "text", "text": "Describe this image briefly."},
+                        {"type": "image_url", "image_url": {"url": create_test_image(color)}},
+                    ],
+                },
+                {
+                    "role": "assistant",
+                    "content": f"This is a {color} colored square image.",
+                },
+            ],
+            "metadata": {"example_id": f"describe_{i}", "type": "description"},
+        })
+    
+    # Write JSONL
+    with output_file.open("w", encoding="utf-8") as f:
+        for example in examples:
+            f.write(json.dumps(example) + "\n")
+    
+    print(f"✅ Created {len(examples)} vision SFT examples")
+    print(f"   Output: {output_file}")
+    print(f"   Size: {output_file.stat().st_size / 1024:.1f} KB")
+    
+    # Validate with SDK
+    try:
+        from synth_ai.learning.sft.data import load_jsonl, validate_vision_example
+        
+        loaded = load_jsonl(output_file, min_messages=1)
+        print(f"   Loaded: {len(loaded)} examples")
+        
+        valid_count = 0
+        for ex in loaded:
+            is_valid, error = validate_vision_example(ex, require_images=True)
+            if is_valid:
+                valid_count += 1
+            else:
+                print(f"   ⚠️  Invalid example: {error}")
+        
+        print(f"   Valid: {valid_count}/{len(loaded)}")
+    except ImportError:
+        print("   (SDK validation skipped - synth_ai not available)")
+
+
+if __name__ == "__main__":
+    main()

examples/rl/README.md

@@ -52,7 +52,7 @@ uvx synth-ai serve math-single-step \
   --port 8101 \
   --env-file examples/rl/.env \
   --trace traces/math \
-  --trace-db traces/math/synth_ai.db
+  --trace-db traces/math/task_app_traces_<timestamp>.db
 ```
 
 Deploy or serve on Modal using the same env file; the registration includes a `ModalDeploymentConfig` that installs the `datasets` package automatically.

examples/rl/configs/eval_base_qwen.toml

@@ -0,0 +1,17 @@
+type = "rl"
+
+provider = "synth"
+task_app_url = "http://localhost:8101"
+model = "Qwen/Qwen3-1.7B"
+split = "validation"
+num_episodes = 50
+seed_start = 0
+
+[policy]
+inference_url = "https://agent-learning.onrender.com/api/inference"
+max_tokens = 128
+temperature = 0.0
+
+# Optionally supply custom headers
+# [policy.headers]
+# Authorization = "Bearer ..."

examples/rl/configs/eval_rl_qwen.toml

@@ -0,0 +1,13 @@
+type = "rl"
+
+provider = "synth"
+task_app_url = "https://your-math-task.modal.run"
+model = "rl:REPLACE_WITH_JOB_ID"
+split = "test"
+num_episodes = 200
+seed_start = 100000
+
+[policy]
+inference_url = "https://your-inference-host"
+max_tokens = 128
+temperature = 0.0

examples/rl/configs/rl_from_base_qwen.toml

@@ -0,0 +1,37 @@
+type = "rl"
+
+[services]
+task_url = "https://your-math-task.modal.run"
+
+[model]
+base = "Qwen/Qwen3-4B"
+
+[policy]
+model = "Qwen/Qwen3-4B"
+inference_url = "https://your-inference-host"
+max_tokens = 128
+temperature = 0.0
+
+[data]
+split = "train"
+seed_start = 0
+episodes_per_iteration = 2048
+evaluation_split = "validation"
+evaluation_episodes = 256
+
+[training]
+max_turns = 1
+ops = ["agent", "env"]
+batch_size = 128
+group_size = 1024
+reward_positive = 1.0
+reward_negative_no_tool = -1.0
+reward_negative_no_answer = -0.5
+learning_rate = 5e-6
+
+[compute]
+gpu_type = "A10G"
+gpu_count = 4
+
+[tags]
+experiment = "math_single_step"

examples/rl/configs/rl_from_base_qwen17.toml

@@ -0,0 +1,76 @@
+type = "rl"
+
+[algorithm]
+type = "online"
+method = "policy_gradient"
+variety = "gspo"
+
+[services]
+task_url = "http://localhost:8101"
+
+[model]
+base = "Qwen/Qwen3-1.7B"
+
+[policy]
+model = "Qwen/Qwen3-1.7B"
+inference_url = "https://agent-learning.onrender.com/api/inference"
+max_tokens = 1028
+temperature = 0.2
+
+[data]
+split = "train"
+seed_start = 0
+episodes_per_iteration = 1280  # 8 per group * 4 groups per batch * 2 batches per step * 20 steps
+evaluation_split = "validation"
+evaluation_episodes = 50
+
+[training]
+max_turns = 1
+ops = ["agent", "env"]
+batch_size = 2
+group_size = 16
+reward_positive = 1.0
+reward_negative_no_tool = -1.0
+reward_negative_no_answer = -0.5
+learning_rate = 5e-6
+log_interval = 1
+weight_sync_interval = 1
+
+[training.weight_sync]
+enable = true
+targets = ["policy"]
+
+[compute]
+gpu_type = "H100"
+gpu_count = 4
+
+[topology]
+type = "single_node_split"
+gpus_for_vllm = 2
+gpus_for_training = 1
+gpus_for_ref = 1
+tensor_parallel = 1
+
+[vllm]
+tensor_parallel_size = 1
+max_model_len = 4096
+
+[reference]
+placement = "dedicated"
+port = 8002
+tp = 1
+health_max_wait_s = 180
+health_interval_ms = 300
+
+[rollout]
+policy_name = "math-single-step"
+max_turns = 1
+episodes_per_batch = 32  # group_size * batch_size
+
+[evaluation]
+instances = 32
+every_n_iters = 10
+seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
+
+[tags]
+experiment = "math_single_step_qwen17"

examples/rl/configs/rl_from_ft_qwen.toml

@@ -0,0 +1,37 @@
+type = "rl"
+
+[services]
+task_url = "https://your-math-task.modal.run"
+
+[model]
+source = "ft:REPLACE_WITH_MODEL_ID"
+
+[policy]
+model = "ft:REPLACE_WITH_MODEL_ID"
+inference_url = "https://your-inference-host"
+max_tokens = 128
+temperature = 0.0
+
+[data]
+split = "train"
+seed_start = 0
+episodes_per_iteration = 2048
+evaluation_split = "validation"
+evaluation_episodes = 256
+
+[training]
+max_turns = 1
+ops = ["agent", "env"]
+batch_size = 128
+group_size = 1024
+reward_positive = 1.0
+reward_negative_no_tool = -1.0
+reward_negative_no_answer = -0.5
+learning_rate = 5e-6
+
+[compute]
+gpu_type = "A10G"
+gpu_count = 4
+
+[tags]
+experiment = "math_single_step_from_fft"