synth-ai 0.2.9.dev3__py3-none-any.whl → 0.2.9.dev4__py3-none-any.whl

examples/warming_up_to_rl/run_eval.py

@@ -0,0 +1,434 @@
+#!/usr/bin/env python3
+"""
+Baseline evaluation script (public-friendly skeleton)
+- Targets a task app (Crafter-like) via initialize/step/terminate
+- Uses a TaskAppClient interface (to be implemented in synth-ai SDK)
+- Keeps structure aligned with research/testing/crafter eval harness
+"""
+from __future__ import annotations
+import os
+import json
+import re
+from typing import Any, Dict, List, Optional
+from collections import Counter
+import asyncio
+import httpx
+import argparse
+import tomllib
+from pathlib import Path
+
+class TaskAppClient:
+    """Minimal async client for the task app initialize/step/terminate routes.
+
+    This is a public-friendly shim for examples, pending SDK surface consolidation.
+    """
+
+    def __init__(self, base_url: str, api_key: Optional[str] = None) -> None:
+        self.base_url = base_url.rstrip("/")
+        self.api_key = api_key
+        self._client: Optional[httpx.AsyncClient] = None
+
+    async def __aenter__(self) -> "TaskAppClient":
+        headers = {}
+        if self.api_key:
+            headers["X-API-Key"] = self.api_key
+        self._client = httpx.AsyncClient(
+            base_url=self.base_url, headers=headers, timeout=600.0, follow_redirects=True
+        )
+        return self
+
+    async def __aexit__(self, exc_type, exc, tb) -> None:
+        if self._client is not None:
+            await self._client.aclose()
+            self._client = None
+
+    @property
+    def client(self) -> httpx.AsyncClient:
+        if self._client is None:
+            # Fallback for direct use without context manager
+            headers = {}
+            if self.api_key:
+                headers["X-API-Key"] = self.api_key
+            self._client = httpx.AsyncClient(
+                base_url=self.base_url, headers=headers, timeout=600.0, follow_redirects=True
+            )
+        return self._client
+
+    async def initialize(self, env_name: str, config: Dict[str, Any]) -> Dict[str, Any]:
+        """POST /env/{env_name}/initialize (compat route supported in task app)."""
+        payload: Dict[str, Any] = {
+            "seed": config.get("seed"),
+        }
+        # Allow both world_config and config inputs; env routes will normalize difficulty
+        if "world_config" in config:
+            payload["world_config"] = config["world_config"]
+        if "config" in config:
+            payload["config"] = config["config"]
+        resp = await self.client.post(f"/env/{env_name}/initialize", json=payload)
+        resp.raise_for_status()
+        return resp.json()
+
+    async def step(self, env_name: str, env_id: str, tool_calls: List[Dict[str, Any]]) -> Dict[str, Any]:
+        """POST /env/{env_name}/step with wrapped tool_calls in action."""
+        payload = {"env_id": env_id, "action": {"tool_calls": tool_calls}}
+        resp = await self.client.post(f"/env/{env_name}/step", json=payload)
+        resp.raise_for_status()
+        return resp.json()
+
+    async def terminate(self, env_name: str, env_id: str) -> Dict[str, Any]:
+        resp = await self.client.post(f"/env/{env_name}/terminate", json={"env_id": env_id})
+        resp.raise_for_status()
+        return resp.json()
+
+    async def get_info(self) -> Dict[str, Any]:
+        resp = await self.client.get("/info")
+        resp.raise_for_status()
+        return resp.json()
+
+    async def proxy_groq_chat(self, payload: Dict[str, Any]) -> Dict[str, Any]:
+        resp = await self.client.post("/proxy/groq/v1/chat/completions", json=payload)
+        resp.raise_for_status()
+        return resp.json()
+
+    async def vllm_chat(self, vllm_base_url: str, payload: Dict[str, Any]) -> Dict[str, Any]:
+        async with httpx.AsyncClient(base_url=vllm_base_url.rstrip("/"), timeout=60.0) as c:
+            resp = await c.post("/v1/chat/completions", json=payload)
+            # Do not raise for status to surface body in errors
+            try:
+                data = resp.json()
+            except Exception:
+                data = {"error": "invalid_json", "raw": resp.text[:800]}
+            if resp.status_code >= 400:
+                return {"error": data}
+            return data
+
+    async def rollout(self, *, run_id: str, env_name: str, seed: int, difficulty: str, policy_name: str, policy_config: Dict[str, Any], max_turns: int) -> Dict[str, Any]:
+        ops: List[str] = []
+        for _ in range(max_turns):
+            ops.extend(["agent", "env"])
+        payload: Dict[str, Any] = {
+            "run_id": run_id,
+            "env": {
+                "env_name": env_name,
+                "config": {"difficulty": difficulty},
+                "seed": seed,
+            },
+            "policy": {
+                "policy_name": policy_name,
+                "config": policy_config,
+            },
+            "ops": ops,
+            "on_done": "terminate",
+        }
+        # Ensure X-API-Key is included
+        headers = {}
+        if self.api_key:
+            headers["X-API-Key"] = self.api_key
+        resp = await self.client.post("/rollout", json=payload, headers=headers)
+        resp.raise_for_status()
+        return resp.json()
+
+TASK_APP_URL = os.getenv("TASK_APP_URL", "https://YOUR-TASK-APP.modal.run").rstrip("/")
+MODEL = os.getenv("EVAL_MODEL", "qwen/qwen3-32b")
+NUM_EPISODES = int(os.getenv("NUM_EPISODES", "3"))
+MAX_TURNS = int(os.getenv("MAX_TURNS", "10"))
+CONCURRENCY = int(os.getenv("CONCURRENCY", "1"))
+
+def _interact_tool_schema() -> List[Dict[str, Any]]:
+    return [{
+        "type": "function",
+        "function": {
+            "name": "interact",
+            "description": "Perform actions in the Crafter environment.",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "actions": {"type": "array", "items": {"type": "string"}},
+                    "reasoning": {"type": "string"},
+                },
+                "required": ["actions", "reasoning"],
+            },
+        },
+    }]
+
+def _build_messages_from_observation(observation: Dict[str, Any], history: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+    inv = observation.get("inventory") or {}
+    pos = observation.get("player_position") or []
+    ach = observation.get("achievements_status") or {}
+    turns_taken = observation.get("num_steps_taken") or 0
+    user_lines: List[str] = []
+    user_lines.append("Environment: CrafterClassic")
+    user_lines.append(f"Player position: {pos}")
+    user_lines.append(f"Inventory: {json.dumps(inv, ensure_ascii=False)}")
+    unlocked = [k for k, v in ach.items() if v]
+    if unlocked:
+        user_lines.append(f"Unlocked achievements: {unlocked}")
+    user_lines.append("Provide 2-5 actions as a plan to explore and progress.")
+    # short history summary
+    if history:
+        last = history[-1]
+        user_lines.append(f"Last actions: {last.get('actions')}")
+    content = "\n".join(user_lines)
+    return [{"role": "user", "content": content}]
+
+def _parse_tool_calls_from_openai_response(data: Dict[str, Any]) -> List[str]:
+    try:
+        choices = data.get("choices")
+        if isinstance(choices, list) and choices:
+            msg = choices[0].get("message", {}) if isinstance(choices[0], dict) else {}
+            tcs = msg.get("tool_calls")
+            if isinstance(tcs, list) and tcs:
+                fn = tcs[0].get("function", {}) if isinstance(tcs[0], dict) else {}
+                args = fn.get("arguments")
+                if isinstance(args, str):
+                    try:
+                        obj = json.loads(args)
+                    except Exception:
+                        obj = {}
+                elif isinstance(args, dict):
+                    obj = args
+                else:
+                    obj = {}
+                acts = obj.get("actions")
+                if isinstance(acts, list):
+                    return [str(a) for a in acts][:5]
+    except Exception:
+        pass
+    # Fallback: parse JSON object from assistant text
+    try:
+        choices = data.get("choices")
+        msg = choices[0].get("message", {}) if isinstance(choices, list) and choices else {}
+        content = msg.get("content")
+        text = ""
+        if isinstance(content, str):
+            text = content
+        elif isinstance(content, list):
+            text = "\n".join(str(part.get("text")) for part in content if isinstance(part, dict) and part.get("text"))
+        for raw in re.findall(r"\{[\s\S]*\}", text or ""):
+            try:
+                obj = json.loads(raw)
+                if isinstance(obj, dict) and obj.get("tool") in ("interact", None):
+                    acts = obj.get("args", {}).get("actions")
+                    if isinstance(acts, list):
+                        return [str(a) for a in acts][:5]
+            except Exception:
+                continue
+    except Exception:
+        pass
+    return []
+
+async def _choose_actions_via_llm(client: TaskAppClient, provider: str, model: str, observation: Dict[str, Any], history: List[Dict[str, Any]]) -> List[str]:
+    messages = _build_messages_from_observation(observation, history)
+    payload: Dict[str, Any] = {
+        "model": model,
+        "messages": messages,
+        "tools": _interact_tool_schema(),
+        "max_tokens": 8192,
+        "temperature": 0.2,
+    }
+    if provider == "groq":
+        # Groq path: avoid forcing tool_choice to reduce 400 errors; proxy will synthesize if missing
+        data = await client.proxy_groq_chat(payload)
+    elif provider == "vllm":
+        info = await client.get_info()
+        vllm_base = ((info or {}).get("inference") or {}).get("base_url")
+        if not vllm_base:
+            return []
+        # For vLLM path, we can force the single tool
+        vllm_payload = dict(payload)
+        vllm_payload["tool_choice"] = {"type": "function", "function": {"name": "interact"}}
+        data = await client.vllm_chat(str(vllm_base), vllm_payload)
+        if isinstance(data, dict) and data.get("error"):
+            return []
+    else:
+        return []
+    actions = _parse_tool_calls_from_openai_response(data)
+    return actions or []
+
+def _expand_actions_to_tool_calls(actions: List[str]) -> List[Dict[str, Any]]:
+    out: List[Dict[str, Any]] = []
+    for a in actions[:5]:
+        out.append({"tool": "interact", "args": {"action": a}})
+    return out
+
+def _detect_provider(model: str) -> str:
+    m = (model or "").lower()
+    if "qwen/qwen3-32b" in m or "qwen-2.5-" in m or m.startswith("groq:"):
+        return "groq"
+    return "vllm"
+
+def _rollout_inference_url_from_cfg(cfg: Dict[str, Any], default_vllm: Optional[str]) -> Optional[str]:
+    # Prefer explicit inference_url in TOML; else fall back to discovered vLLM base
+    url = cfg.get("inference_url")
+    if isinstance(url, str) and url:
+        return url
+    return default_vllm
+
+async def eval_episode(client: TaskAppClient, seed: int) -> Dict[str, Any]:
+    env_name = "CrafterClassic"
+    history: List[Dict[str, Any]] = []
+    achievements: set[str] = set()
+    turns = 0
+
+    # Initialize environment
+    init_cfg: Dict[str, Any] = {"seed": seed, "world_config": {"difficulty": os.getenv("DIFFICULTY", "easy")}}
+    created = await client.initialize(env_name, init_cfg)
+    env_id = created.get("env_id")
+    if not isinstance(env_id, str) or not env_id:
+        raise RuntimeError(f"Invalid env_id from initialize: {created}")
+    done = False
+    provider = _detect_provider(MODEL)
+    observation = created.get("observation") if isinstance(created, dict) else None
+    if not isinstance(observation, dict):
+        observation = {}
+
+    try:
+        while turns < MAX_TURNS and not done:
+            # Ask LLM for actions; fallback to a simple exploratory pair
+            chosen_actions = await _choose_actions_via_llm(client, provider, MODEL, observation, history)
+            if not chosen_actions:
+                chosen_actions = ["move_up", "do"]
+            tool_calls = _expand_actions_to_tool_calls(chosen_actions)
+            step = await client.step(env_name, env_id, tool_calls)
+            done = bool(step.get("done"))
+            turns += 1
+            history.append({"actions": chosen_actions, "reasoning": "explore then interact"})
+            # Update observation for next turn if available
+            if isinstance(step, dict):
+                nxt = step.get("observation")
+                if isinstance(nxt, dict):
+                    observation = nxt
+    finally:
+        try:
+            await client.terminate(env_name, env_id)
+        except Exception:
+            pass
+
+    return {"seed": seed, "turns": turns, "achievements": sorted(achievements)}
+
+async def main() -> None:
+    # Best-effort load local .env if present (ensures ENVIRONMENT_API_KEY for rollout)
+    try:
+        env_path = Path(__file__).resolve().parent / ".env"
+        if env_path.exists():
+            for line in env_path.read_text(encoding="utf-8").splitlines():
+                line = line.strip()
+                if not line or line.startswith("#") or "=" not in line:
+                    continue
+                k, v = line.split("=", 1)
+                k = k.strip()
+                v = v.strip().strip('"').strip("'")
+                os.environ.setdefault(k, v)
+    except Exception:
+        pass
+
+    parser = argparse.ArgumentParser(description="Baseline eval against task app with optional TOML config")
+    parser.add_argument("--toml", help="Path to TOML config file", default=None)
+    parser.add_argument("--use-rollout", action="store_true", help="Use server-side rollout endpoint for eval")
+    args = parser.parse_args()
+
+    global TASK_APP_URL, MODEL, NUM_EPISODES, MAX_TURNS, CONCURRENCY
+    cfg: Dict[str, Any] = {}
+    if args.toml:
+        with open(args.toml, "rb") as f:
+            cfg = tomllib.load(f)
+        # Map known keys; tolerate missing
+        TASK_APP_URL = (cfg.get("task_app_url") or TASK_APP_URL).rstrip("/")
+        MODEL = cfg.get("model") or MODEL
+        NUM_EPISODES = int(cfg.get("num_episodes") or NUM_EPISODES)
+        MAX_TURNS = int(cfg.get("max_turns") or MAX_TURNS)
+        CONCURRENCY = int(cfg.get("concurrency") or CONCURRENCY)
+        if "difficulty" in cfg:
+            os.environ["DIFFICULTY"] = str(cfg.get("difficulty"))
+        # Replace placeholder URLs with env if present
+        if "your-task-app.modal.run" in TASK_APP_URL.lower():
+            env_url = os.getenv("TASK_APP_URL")
+            if env_url:
+                TASK_APP_URL = env_url.rstrip("/")
+            else:
+                raise RuntimeError("TASK_APP_URL is a placeholder. Set task_app_url in TOML or export TASK_APP_URL.")
+
+    print(f"Task App: {TASK_APP_URL}")
+    print(f"Model: {MODEL} Episodes: {NUM_EPISODES} Max turns: {MAX_TURNS} Concurrency: {CONCURRENCY}")
+    sem = asyncio.Semaphore(max(CONCURRENCY, 1))
+    async with TaskAppClient(TASK_APP_URL, api_key=os.getenv("ENVIRONMENT_API_KEY")) as client:
+        if args.use_rollout:
+            # Use server-side rollout; derive inference URL per provider
+            info = await client.get_info()
+            default_vllm = ((info or {}).get("inference") or {}).get("base_url")
+            inf_url = _rollout_inference_url_from_cfg(cfg, default_vllm)
+            if not inf_url:
+                raise RuntimeError("Could not resolve inference URL for rollout")
+            async def _run(seed: int):
+                async with sem:
+                    try:
+                        run_id = f"eval-{seed}"
+                        # Build policy config from TOML (explicit control; no server-side guessing)
+                        policy_cfg: Dict[str, Any] = {
+                            "model": cfg.get("model", MODEL),
+                            "inference_url": inf_url,
+                        }
+                        for k in ("max_tokens", "temperature", "top_p", "thinking_mode", "thinking_budget", "use_tools"):
+                            if k in cfg and cfg.get(k) is not None:
+                                policy_cfg[k] = cfg.get(k)
+
+                        r = await client.rollout(
+                            run_id=run_id,
+                            env_name="crafter",
+                            seed=seed,
+                            difficulty=os.getenv("DIFFICULTY", "easy"),
+                            policy_name=cfg.get("policy_name", "crafter"),
+                            policy_config=policy_cfg,
+                            max_turns=MAX_TURNS,
+                        )
+                        # Extract achievements count if present
+                        ach = []
+                        try:
+                            trajs = r.get("trajectories") or []
+                            final_obs = (trajs[0].get("final") or {}).get("observation") if trajs and isinstance(trajs[0], dict) else None
+                            ach_map = (final_obs or {}).get("achievements_status") if isinstance(final_obs, dict) else None
+                            if isinstance(ach_map, dict):
+                                ach = sorted([k for k, v in ach_map.items() if v])
+                        except Exception:
+                            pass
+                        length = 0
+                        try:
+                            trajs = r.get("trajectories") or []
+                            if trajs and isinstance(trajs[0], dict):
+                                length = int(trajs[0].get("length") or 0)
+                        except Exception:
+                            pass
+                        return {"seed": seed, "turns": length, "achievements": ach}
+                    except Exception as e:
+                        return {"seed": seed, "turns": 0, "achievements": [], "error": str(e)}
+            results = await asyncio.gather(*[asyncio.create_task(_run(i)) for i in range(1, NUM_EPISODES + 1)], return_exceptions=False)
+            # Aggregate summary
+            counts = [len(r.get("achievements") or []) for r in results if isinstance(r, dict)]
+            turns = [int(r.get("turns") or 0) for r in results if isinstance(r, dict)]
+            all_ach = Counter()
+            for r in results:
+                try:
+                    for a in r.get("achievements") or []:
+                        all_ach[a] += 1
+                except Exception:
+                    pass
+            summary = {
+                "episodes": results,
+                "aggregate": {
+                    "completed": sum(1 for r in results if not r.get("error")),
+                    "total": len(results),
+                    "avg_turns": (sum(turns) / len(turns)) if turns else 0.0,
+                    "avg_achievements": (sum(counts) / len(counts)) if counts else 0.0,
+                    "achievements_freq": dict(all_ach),
+                },
+            }
+            print(json.dumps(summary, indent=2))
+        else:
+            async def _run(seed: int):
+                async with sem:
+                    return await eval_episode(client, seed)
+            results = await asyncio.gather(*[asyncio.create_task(_run(i)) for i in range(1, NUM_EPISODES + 1)])
+            print(json.dumps({"episodes": results}, indent=2))
+
+if __name__ == "__main__":
+    asyncio.run(main())