synth-ai 0.2.16__py3-none-any.whl → 0.2.19__py3-none-any.whl

examples/blog_posts/warming_up_to_rl/configs/eval_baseline_qwen32b_10x20.toml

@@ -0,0 +1,25 @@
+[eval]
+app_id = "grpo-crafter"
+task_app_url = "https://synth-laboratories--grpo-crafter-task-app-fastapi-app-dev.modal.run"
+model = "Qwen/Qwen3-4B"
+seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
+max_turns = 20
+concurrency = 1
+env_name = "crafter"
+policy_name = "crafter-react"
+trace_format = "structured"
+return_trace = true
+
+[eval.policy_config]
+provider = "synth"
+model = "Qwen/Qwen3-4B"
+inference_url = "https://synth-laboratories-dev--learning-v2-service-fastapi-app.modal.run"
+temperature = 0.6
+top_p = 0.95
+max_tokens = 2048
+use_vision = false
+image_only_mode = false
+max_llm_calls = 10
+
+[eval.env_config.env_params]
+max_steps_per_episode = 20

examples/blog_posts/warming_up_to_rl/configs/eval_ft_qwen4b.toml

@@ -0,0 +1,25 @@
+# Evaluate the finetuned Qwen3-4B checkpoint on Crafter.
+# Replace model with the fft: job id returned by the SFT run.
+
+[eval]
+app_id = "grpo-crafter"
+task_app_url = "https://synth-laboratories--grpo-crafter-task-app-fastapi-app-dev.modal.run"
+model = "fft:REPLACE-WITH-SFT-JOB-ID"
+seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
+max_turns = 10
+concurrency = 4
+env_name = "crafter"
+policy_name = "crafter-react"
+trace_format = "compact"
+return_trace = false
+
+[eval.policy_config]
+provider = "synth"
+model = "fft:REPLACE-WITH-SFT-JOB-ID"
+temperature = 0.2
+top_p = 0.8
+max_tokens = 512
+use_vision = true
+image_only_mode = false
+max_llm_calls = 10
+tool_choice = "auto"

examples/blog_posts/warming_up_to_rl/configs/eval_ft_qwen4b_10x20.toml

@@ -0,0 +1,26 @@
+[eval]
+app_id = "grpo-crafter"
+task_app_url = "https://synth-laboratories--grpo-crafter-task-app-fastapi-app-dev.modal.run"
+model = "peft:Qwen/Qwen3-4B:job_f774218e6c954517"
+seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
+max_turns = 20
+concurrency = 2
+env_name = "crafter"
+policy_name = "crafter-react"
+trace_format = "structured"
+return_trace = true
+
+[eval.policy_config]
+provider = "synth"
+model = "peft:Qwen/Qwen3-4B:job_f774218e6c954517"
+inference_url = "https://synth-laboratories-dev--learning-v2-service-fastapi-app.modal.run"
+temperature = 0.2
+top_p = 0.8
+max_tokens = 1024
+use_vision = false
+image_only_mode = false
+max_llm_calls = 10
+tool_choice = "auto"
+
+[eval.env_config.env_params]
+max_steps_per_episode = 20

examples/blog_posts/warming_up_to_rl/configs/eval_groq_qwen32b.toml

@@ -0,0 +1,25 @@
+[eval]
+app_id = "grpo-crafter"
+task_app_url = "https://synth-laboratories--grpo-crafter-task-app-fastapi-app-dev.modal.run"
+model = "qwen/qwen3-32b"
+seeds = [ 0, 1, 2,]
+max_turns = 10
+concurrency = 1
+env_name = "crafter"
+policy_name = "crafter-react"
+trace_format = "full"
+return_trace = true
+
+[eval.policy_config]
+provider = "groq"
+model = "qwen/qwen3-32b"
+inference_url = "https://api.groq.com/openai"
+temperature = 0.6
+top_p = 0.95
+max_tokens = 8192
+use_vision = false
+image_only_mode = false
+max_llm_calls = 10
+
+[eval.env_config.env_params]
+max_steps_per_episode = 10

examples/blog_posts/warming_up_to_rl/configs/eval_openai_gpt_oss_120b.toml

@@ -0,0 +1,29 @@
+# Crafter rollout config for GPT-OSS-120B served from OpenAI-compatible APIs.
+# Replace the task_app_url with your deployed Crafter task app URL.
+# The run stores full traces so we can keep the LLM reasoning for fine-tuning.
+
+[eval]
+app_id = "grpo-crafter"
+task_app_url = "https://synth-laboratories--grpo-crafter-task-app-fastapi-app-dev.modal.run"
+model = "openai/gpt-oss-120b"
+seeds = [0, 1, 2]
+max_turns = 10
+concurrency = 1
+env_name = "crafter"
+policy_name = "crafter-react"
+trace_format = "full"
+return_trace = true
+
+[eval.env_config]
+env_params = { max_steps_per_episode = 10 }
+
+[eval.policy_config]
+provider = "groq"
+model = "openai/gpt-oss-120b"
+inference_url = "https://api.groq.com/openai"
+temperature = 0.6
+top_p = 0.9
+max_tokens = 768
+use_vision = false
+image_only_mode = false
+max_llm_calls = 10

examples/blog_posts/warming_up_to_rl/configs/filter_high_reward_dataset.toml

@@ -0,0 +1,10 @@
+# Filters Crafter traces into an instruction-tuning dataset.
+# Assumes you stored rollouts in traces/v3/crafter_blog.db via `uvx synth-ai eval`.
+
+[filter]
+db = "sqlite+libsql://http://127.0.0.1:8080"
+output = "examples/blog_posts/warming_up_to_rl/ft_data/crafter_blog_high_reward.jsonl"
+min_official_score = 0.1
+models = ["qwen/qwen3-32b", "openai/gpt-oss-120b"]
+shuffle = true
+shuffle_seed = 42

examples/blog_posts/warming_up_to_rl/configs/smoke_test.toml

@@ -0,0 +1,75 @@
+# Example RL config with smoke testing enabled
+# This config demonstrates auto-starting task app and sqld for easy smoke testing
+
+type = "rl"
+
+# Smoke testing configuration - AUTO-STARTS services in background!
+[smoke]
+# Auto-start the task app server
+task_app_name = "grpo-crafter"  # Your task app name (use "synth-ai task-app list" to see available apps)
+task_app_port = 8765
+task_app_env_file = ".env"  # Required for this task app
+task_app_force = true  # Kill any existing process on this port
+
+# Auto-start sqld for tracing
+sqld_auto_start = true
+sqld_db_path = "./traces/local.db"
+sqld_hrana_port = 8080
+sqld_http_port = 8081
+
+# Test parameters
+env_name = "crafter"
+policy_name = "crafter-react"
+max_steps = 10
+policy = "gpt-5-nano"  # Use gpt-5-nano policy with mock backend
+model = "gpt-4o-mini"  # Real model to use via OpenAI
+mock_backend = "openai"  # Use OpenAI backend for real inference and tool calls
+return_trace = true
+use_mock = true  # Use mock proxy that routes to OpenAI
+
+# RL Training Configuration (used by actual training, not smoke tests)
+[algorithm]
+type = "online"
+method = "policy_gradient"
+variety = "gspo"
+
+[policy]
+model_name = "Qwen/Qwen3-4B"
+trainer_mode = "full"
+label = "crafter-rl-demo"
+
+[compute]
+gpu_type = "H100"
+gpu_count = 2
+
+[compute.topology]
+type = "single_node_split"
+gpus_for_vllm = 1
+gpus_for_training = 1
+
+[services]
+task_url = "http://localhost:8765"
+
+[rollout]
+env_name = "crafter"
+policy_name = "crafter-react"
+max_turns = 10
+episodes_per_batch = 16
+max_concurrent_rollouts = 4
+task_app_origin_rewards_only = true
+
+[training]
+num_epochs = 1
+iterations_per_epoch = 10
+max_turns = 10
+batch_size = 4
+group_size = 4
+learning_rate = 5e-5
+weight_sync_interval = 1
+log_interval = 1
+
+[evaluation]
+instances = 2
+every_n_iters = 1
+seeds = [0, 1]
+

examples/blog_posts/warming_up_to_rl/configs/train_rl_from_sft.toml

@@ -0,0 +1,91 @@
+# Continue training the finetuned Crafter policy with GRPO-style RL.
+# Fill in task_url with your deployed task app and set model.source to the
+# finetuned model id returned by `uvx synth-ai train --type sft`.
+
+type = "rl"
+
+# [smoke] section is OPTIONAL and only used by `synth-ai smoke` command for local testing.
+# This section is completely IGNORED by the RL trainer and will not affect training jobs.
+# It allows you to quickly test your task app without passing many CLI arguments:
+#   uvx synth-ai smoke --config this-file.toml
+# All values are optional; CLI args override TOML values.
+[smoke]
+task_url = "https://synth-laboratories--crafter-blogpost-fastapi-app-dev.modal.run"
+env_name = "crafter"
+policy_name = "crafter-react"
+max_steps = 10
+policy = "mock"  # mock, gpt-5-nano, openai, groq
+model = "gpt-5-nano"
+mock_backend = "openai"  # synthetic or openai
+mock_port = 0  # 0 = auto-assign
+return_trace = true
+use_mock = true
+
+[algorithm]
+type = "online"
+method = "policy_gradient"
+variety = "gspo"
+
+[services]
+task_url = "https://synth-laboratories--crafter-blogpost-fastapi-app-dev.modal.run"
+judge_url = "https://synth-backend-dev-docker.onrender.com/api"
+
+[compute]
+gpu_type = "H200"
+gpu_count = 2
+[compute.topology]
+reference_placement = "none"
+
+[topology]
+type = "single_node_split"
+reference_placement = "none"
+gpus_for_vllm = 1
+gpus_for_training = 1
+gpus_for_ref = 0
+tensor_parallel = 1
+
+[vllm]
+tensor_parallel_size = 1
+max_model_len = 8192
+
+[reference]
+placement = "none"
+
+[model]
+base = "Qwen/Qwen3-4B"
+trainer_mode = "lora"
+label = "crafter-rl-baseline"
+
+[rollout]
+env_name = "crafter"
+policy_name = "crafter-react"
+max_turns = 10
+episodes_per_batch = 20
+max_concurrent_rollouts = 8
+rubric_rewards_only = false
+task_app_origin_rewards_only = true
+
+[evaluation]
+instances = 100
+every_n_iters = 20
+seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99]
+
+[training]
+num_epochs = 1
+iterations_per_epoch = 1
+max_turns = 10
+batch_size = 2
+group_size = 2
+learning_rate = 5e-6
+weight_sync_interval = 1
+log_interval = 1
+max_completion_tokens = 256
+async_semaphore_max = 4
+
+[training.weight_sync]
+enable = true
+targets = ["policy"]
+weight_sync_interval = 1
+
+[rubric]
+enabled = false

examples/blog_posts/warming_up_to_rl/configs/train_sft_qwen4b.toml

@@ -0,0 +1,40 @@
+# Full-finetune Qwen3-4B on filtered Crafter traces.
+# Update the `data` path once `uvx synth-ai filter` produces your JSONL.
+
+[algorithm]
+type = "offline"
+method = "sft"
+variety = "fft"
+
+[job]
+model = "Qwen/Qwen3-4B"
+data = "examples/blog_posts/warming_up_to_rl/ft_data/crafter_blog_high_reward.jsonl"
+poll_seconds = 1800
+
+[compute]
+gpu_type = "H100"
+gpu_count = 4
+nodes = 1
+
+[data.topology]
+container_count = 4
+
+[training]
+mode = "full_finetune"
+use_qlora = false
+
+[hyperparameters]
+n_epochs = 2
+world_size = 4
+sequence_length = 2048
+per_device_batch = 2
+gradient_accumulation_steps = 64
+learning_rate = 8e-6
+warmup_ratio = 0.03
+
+[hyperparameters.parallelism]
+use_deepspeed = true
+deepspeed_stage = 3
+fsdp = false
+bf16 = true
+fp16 = false

examples/blog_posts/warming_up_to_rl/warming_up_to_rl_baseline.py

@@ -0,0 +1,187 @@
+"""Warming Up to RL baseline for Crafter.
+
+This baseline demonstrates how to evaluate an LLM agent on the Crafter survival game
+without requiring a deployed task app. This is the recommended starting point for coding
+agents to get a baseline score before making changes.
+
+Quick Start:
+    # Run a quick 3-task baseline
+    uvx synth-ai baseline warming_up_to_rl --split train --seeds 0,1,2
+
+    # Full train evaluation
+    uvx synth-ai baseline warming_up_to_rl --split train
+
+    # Compare models
+    uvx synth-ai baseline warming_up_to_rl --model groq:openai/gpt-oss-20b
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+try:
+    import crafter
+    CRAFTER_AVAILABLE = True
+except ImportError:
+    CRAFTER_AVAILABLE = False
+
+from synth_ai.baseline import BaselineConfig, BaselineTaskRunner, DataSplit, TaskResult
+from synth_ai.types import EventReward, OutcomeReward
+
+
+class CrafterRunner(BaselineTaskRunner):
+    """Task runner for Crafter environment."""
+
+    def __init__(self, policy_config: dict[str, Any], env_config: dict[str, Any]):
+        super().__init__(policy_config, env_config)
+        self.max_steps = env_config.get("max_steps", 1000)
+
+    async def run_task(self, seed: int) -> TaskResult:
+        """Run a single Crafter episode."""
+        if not CRAFTER_AVAILABLE:
+            raise ImportError(
+                "Crafter not installed. Install with: pip install crafter"
+            )
+
+        # Create environment
+        env = crafter.Env()
+        env.reset()
+
+        # Initialize tracking
+        event_rewards: list[EventReward] = []
+        achievements = {}
+        step_count = 0
+
+        # Get model configuration
+        from synth_ai.inference.client import InferenceClient
+
+        client = InferenceClient()
+        model = self.policy_config.get("model", "gpt-4o-mini")
+        temperature = self.policy_config.get("temperature", 0.7)
+
+        # Define action tool
+        actions = [
+            "noop", "move_left", "move_right", "move_up", "move_down",
+            "do", "sleep", "place_stone", "place_table", "place_furnace",
+            "place_plant", "make_wood_pickaxe", "make_stone_pickaxe",
+            "make_iron_pickaxe", "make_wood_sword", "make_stone_sword",
+            "make_iron_sword"
+        ]
+
+        action_tool = {
+            "type": "function",
+            "function": {
+                "name": "take_action",
+                "description": "Take an action in the Crafter world",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "action": {
+                            "type": "string",
+                            "enum": actions,
+                            "description": f"Action to take. Available: {', '.join(actions)}",
+                        }
+                    },
+                    "required": ["action"],
+                },
+            },
+        }
+
+        # Run episode
+        done = False
+        while not done and step_count < self.max_steps:
+            # Get observation (would include visual state in full implementation)
+            obs_str = f"Crafter Step {step_count}\n"
+            obs_str += f"Current achievements: {achievements}\n"
+            obs_str += "What action should you take to survive and progress?"
+
+            # Get action from model
+            try:
+                response = await client.generate(
+                    model=model,
+                    messages=[
+                        {
+                            "role": "system",
+                            "content": "You are an expert at survival games. Use the take_action tool to survive and achieve goals in Crafter.",
+                        },
+                        {"role": "user", "content": obs_str},
+                    ],
+                    tools=[action_tool],
+                    temperature=temperature,
+                    max_tokens=100,
+                )
+
+                # Extract action
+                action_name = "noop"
+                if response.get("tool_calls"):
+                    tool_call = response["tool_calls"][0]
+                    args = json.loads(tool_call["function"]["arguments"])
+                    action_name = args.get("action", "noop")
+
+                action_idx = actions.index(action_name) if action_name in actions else 0
+
+                # Take step
+                obs, reward, done, info = env.step(action_idx)
+
+                # Update achievements
+                if "achievements" in info:
+                    achievements.update(info["achievements"])
+
+                # Track rewards
+                if reward > 0:
+                    event_rewards.append(
+                        EventReward(
+                            event_id=f"step_{step_count}",
+                            reward=reward,
+                            metadata={"action": action_name, "achievements": achievements.copy()},
+                        )
+                    )
+
+                step_count += 1
+
+            except Exception as e:
+                done = True
+                break
+
+        # Calculate outcome reward based on achievements
+        total_achievements = sum(achievements.values())
+        success = total_achievements >= 3  # At least 3 achievements
+
+        return TaskResult(
+            success=success,
+            outcome_reward=OutcomeReward(
+                reward=float(total_achievements),
+                metadata={
+                    "steps": step_count,
+                    "achievements": achievements,
+                    "seed": seed,
+                },
+            ),
+            event_rewards=event_rewards,
+            total_steps=step_count,
+            metadata={"achievements": achievements},
+        )
+
+
+# Define baseline configuration (only if Crafter is available)
+if CRAFTER_AVAILABLE:
+    warming_up_to_rl_baseline = BaselineConfig(
+        baseline_id="warming_up_to_rl",
+        name="Warming Up to RL - Crafter",
+        description="Crafter survival game baseline for comparing agent performance on RL tasks",
+        task_runner=CrafterRunner,
+        splits={
+            "train": DataSplit(name="train", seeds=list(range(20))),
+            "val": DataSplit(name="val", seeds=list(range(20, 25))),
+            "test": DataSplit(name="test", seeds=list(range(25, 30))),
+        },
+        default_policy_config={
+            "model": "gpt-4o-mini",
+            "temperature": 0.7,
+        },
+        default_env_config={
+            "max_steps": 1000,
+        },
+        tags=["rl", "survival", "achievements", "blog-post"],
+    )

examples/dev/qwen3_32b_qlora_4xh100.toml

@@ -1,3 +1,8 @@
+[algorithm]
+type = "offline"
+method = "sft"
+variety = "qlora"
+
 [job]
 model = "Qwen/Qwen3-32B"
 # Optionally set here; you can also pass --dataset

examples/multi_step/configs/VERILOG_REWARDS.md

@@ -88,3 +88,7 @@ Expected output for successful rollout:
 - `mean_return` ≈ 1.0+ (if full submit success)
 
 
+
+
+
+

examples/multi_step/configs/VERILOG_RL_CHECKLIST.md

@@ -181,3 +181,7 @@ Before starting RL training, verify:
 - [verilog_rl_lora.toml](./verilog_rl_lora.toml) - Training configuration
 
 
+
+
+
+

examples/multi_step/configs/crafter_rl_outcome.toml

@@ -6,7 +6,7 @@ method = "policy_gradient"
 variety = "gspo"
 
 [services]
-# Replace with the Modal URL printed by `uvx synth-ai modal-serve grpo-crafter`
+# Replace with the Modal URL printed by `uvx synth-ai deploy --runtime modal --modal-mode serve grpo-crafter`
 task_url = "https://YOUR-MODAL-TASK-APP.modal.run"
 
 [compute]
@@ -46,6 +46,7 @@ policy_name = "crafter-react"
 max_concurrent_rollouts = 12
 batches_per_step = 2
 ops = ["agent", "env"]
+task_app_origin_rewards_only = true
 
 [evaluation]
 instances = 10