synth-ai 0.2.14__py3-none-any.whl → 0.2.16__py3-none-any.whl

examples/task_apps/crafter/task_app/synth_envs_hosted/policy_routes.py

@@ -417,8 +417,6 @@ async def step_policy(
             inf_req = meta["inference_request"]
             msgs = inf_req["messages"]
             model_name = inf_req.get("model") or getattr(policy, "model", None) or ""
-            system_messages: list[str] = []
-            user_messages: list[str] = []
             if msgs and len(msgs) > 0 and msgs[0]["role"] == "system":
                 sys_text = msgs[0]["content"]
                 policy_name = getattr(policy, "name", "") or type(policy).__name__.lower()
@@ -901,38 +899,71 @@ async def step_policy(
                             req_body["temperature"] = 0.1
                         meta["inference_request"] = req_body
 
-                # Strip image parts: Crafter policy currently only uses text prompts.
-                # Some providers reject image_url payloads entirely, so always flatten to plain text.
-                req_body2 = meta.get("inference_request", {})
-                if isinstance(req_body2, dict):
-                    msgs = req_body2.get("messages")
-                    if isinstance(msgs, list):
-                        new_msgs = []
-                        changed = False
-                        for m in msgs:
-                            try:
-                                if isinstance(m, dict):
-                                    content = m.get("content")
-                                    if isinstance(content, list):
-                                        parts: list[str] = []
-                                        for seg in content:
-                                            if isinstance(seg, dict):
-                                                txt = seg.get("text") or seg.get("content")
-                                                if isinstance(txt, str) and txt:
-                                                    parts.append(txt)
-                                        m2 = dict(m)
-                                        m2["content"] = "\n".join(parts)
-                                        new_msgs.append(m2)
-                                        changed = True
+                # Message flattening: Convert multimodal content to text-only for non-vision models.
+                # SKIP message flattening for vision models to preserve image_url parts!
+                # The old code here was flattening multimodal content (list) to text-only (str),
+                # which strips out image_url parts. This breaks vision models.
+                # Only flatten for non-vision models that can't handle multimodal format.
+                is_vision_model = False
+                try:
+                    # Check if the policy is a vision-capable policy
+                    if isinstance(policy, CrafterPolicy):
+                        is_vision_model = getattr(policy, "use_vision", False)
+                except Exception:
+                    pass
+                
+                logger.debug(f"🔊 [POLICY_ROUTES] is_vision_model={is_vision_model}, will_flatten={not is_vision_model}")
+                
+                if not is_vision_model:
+                    # Only flatten for non-vision models (backward compatibility)
+                    req_body2 = meta.get("inference_request", {})
+                    if isinstance(req_body2, dict):
+                        msgs = req_body2.get("messages")
+                        if isinstance(msgs, list):
+                            new_msgs = []
+                            changed = False
+                            for m in msgs:
+                                try:
+                                    if isinstance(m, dict):
+                                        content = m.get("content")
+                                        if isinstance(content, list):
+                                            parts: list[str] = []
+                                            for seg in content:
+                                                if isinstance(seg, dict):
+                                                    txt = seg.get("text") or seg.get("content")
+                                                    if isinstance(txt, str) and txt:
+                                                        parts.append(txt)
+                                            m2 = dict(m)
+                                            m2["content"] = "\n".join(parts)
+                                            new_msgs.append(m2)
+                                            changed = True
+                                        else:
+                                            new_msgs.append(m)
                                     else:
                                         new_msgs.append(m)
-                                else:
+                                except Exception:
                                     new_msgs.append(m)
-                            except Exception:
-                                new_msgs.append(m)
-                        if changed:
-                            req_body2["messages"] = new_msgs
-                            meta["inference_request"] = req_body2
+                            if changed:
+                                req_body2["messages"] = new_msgs
+                                meta["inference_request"] = req_body2
+                                logger.debug(f"🔊 [POLICY_ROUTES] Flattened messages for non-vision model")
+                else:
+                    logger.debug(f"🔊 [POLICY_ROUTES] Preserving multimodal content for vision model")
+                
+                # DEBUG: Log final message structure before calling inference
+                final_req = meta.get("inference_request", {})
+                if isinstance(final_req, dict):
+                    final_msgs = final_req.get("messages", [])
+                    logger.debug(f"🔊 [POLICY_ROUTES_FINAL] Sending {len(final_msgs)} messages to inference")
+                    for idx, msg in enumerate(final_msgs):
+                        if isinstance(msg, dict):
+                            content = msg.get("content")
+                            logger.debug(f"🔊 [POLICY_ROUTES_FINAL] Message[{idx}]: type={type(content).__name__}, is_list={isinstance(content, list)}")
+                            if isinstance(content, list):
+                                logger.debug(f"🔊 [POLICY_ROUTES_FINAL]   Content list has {len(content)} items")
+                                for part_idx, part in enumerate(content[:3]):  # Show first 3 items
+                                    if isinstance(part, dict):
+                                        logger.debug(f"🔊 [POLICY_ROUTES_FINAL]     Part[{part_idx}]: type={part.get('type')}")
 
             _t_start = _t.time()
             call_started_at = datetime.utcnow()

examples/task_apps/crafter/task_app/synth_envs_hosted/rollout.py

@@ -985,10 +985,15 @@ class RolloutTracingContext:
     def build_trace_payload(self, session_trace: Any) -> dict[str, Any] | None:
         if not self.return_trace or session_trace is None:
             return None
-        if self.trace_format == "full":
+        
+        # For both "full" and "structured" formats, return the complete session trace
+        # The CLI (synth-ai eval) expects this for proper trace storage
+        if self.trace_format in ("full", "structured"):
             payload = session_trace.to_dict()
             payload.setdefault("metadata", {}).update(self.metadata_updates)
             return payload
+        
+        # For "compact" format, return only summary stats
         metadata = dict(session_trace.metadata)
         metadata.update(self.metadata_updates)
         return {
@@ -1173,14 +1178,6 @@ async def execute_rollout(
             logger.debug(f"TRACER_FACTORY_FAIL: {exc}")
     tracing_context = RolloutTracingContext(tracer_instance, request, req)
     await tracing_context.start_session()
-    # Print whether tracing is active for this rollout
-    try:
-        print(
-            f"[rollout] tracing enabled={bool(tracing_context.enabled)} run_id={request.run_id}",
-            flush=True,
-        )
-    except Exception:
-        pass
 
     # Register run
     registry.register_run(request.run_id)
@@ -1625,16 +1622,21 @@ async def execute_rollout(
 
             elif op == "env":
                 if not pending_tool_calls:
+                    # Instead of failing, inject a no-op action to keep the rollout going
                     with contextlib.suppress(Exception):
                         logger.warning(
-                            "POLICY_STEP_FAIL: missing tool_calls; failing rollout run_id=%s op_idx=%s",
+                            "POLICY_STEP_NOOP: missing tool_calls; injecting noop action run_id=%s op_idx=%s",
                             request.run_id,
                             str(op_idx),
                         )
-                    raise HTTPException(
-                        status_code=500,
-                        detail="policy_step_failed: missing tool_calls (no_tool_calls)",
-                    )
+                    # Create a noop tool call in the format expected by the environment
+                    pending_tool_calls = [
+                        {
+                            "id": f"noop_{op_idx}",
+                            "tool": "interact",
+                            "arguments": {"action": "noop"},
+                        }
+                    ]
 
                 # Environment step
                 from .environment_routes import EnvStepRequest, step_environment

examples/task_apps/enron/__init__.py

@@ -1 +1,2 @@
 
+

examples/vlm/README.md

@@ -21,8 +21,8 @@ plumbing with lightweight utilities for dataset curation and training.
 3. **Export multimodal SFT rows**  
    ```
    uv run python examples/warming_up_to_rl/export_trace_sft.py \
-       --db traces/v3/synth_ai.db \
-       --output examples/vlm/output/crafter_traces_full.jsonl
+      --db traces/v3/task_app_traces_<timestamp>.db \
+       --output examples/vlm/output/crafter_sft_full.jsonl
    ```
    The exporter now emits `metadata.has_image`, `metadata.user_has_image`, and
    `metadata.assistant_has_image` flags per turn.
@@ -30,7 +30,7 @@ plumbing with lightweight utilities for dataset curation and training.
 4. **Filter to image-rich turns**  
    ```
    uv run python examples/vlm/filter_image_rows.py \
-       --input examples/vlm/output/crafter_traces_full.jsonl \
+       --input examples/vlm/output/crafter_sft_full.jsonl \
        --output examples/vlm/output/crafter_vlm_dataset.jsonl
    ```
 

examples/vlm/configs/crafter_vlm_gpt4o.toml

@@ -1,3 +1,5 @@
+type = "sft"
+
 [job]
 model = "openai/gpt-4o-mini-2024-07-18"
 modalities = ["text", "image"]

examples/vlm/crafter_openai_vlm_agent.py

@@ -24,6 +24,7 @@ import asyncio
 import base64
 import json
 import os
+from contextlib import suppress
 from pathlib import Path
 from typing import Any
 from uuid import uuid4
@@ -62,7 +63,7 @@ class EpisodeResult:
                 if unlocked:
                     self.achievements.add(str(name))
         reward = obs.get("reward_last_step")
-        if isinstance(reward, (int, float)):
+        if isinstance(reward, int | float):
             self.total_reward += float(reward)
 
 
@@ -107,11 +108,8 @@ def _decode_and_save_image(observation: dict[str, Any], path: Path) -> None:
     if not isinstance(base64_data, str) or not base64_data:
         return
     path.parent.mkdir(parents=True, exist_ok=True)
-    try:
+    with suppress(Exception):
         path.write_bytes(base64.b64decode(base64_data))
-    except Exception:
-        # Best-effort; corrupted frames should not halt rollout
-        pass
 
 
 def _normalise_openai_request(payload: dict[str, Any], model: str, temperature: float) -> dict[str, Any]:

examples/vlm/filter_image_rows.py

@@ -8,7 +8,7 @@ output now that each record's metadata includes `has_image`, `user_has_image`, a
 
 Usage:
     uv run python examples/vlm/filter_image_rows.py \
-        --input examples/sft/ft_data/crafter_traces.jsonl \
+        --input examples/sft/ft_data/crafter_sft.jsonl \
         --output examples/vlm/output/crafter_vlm_dataset.jsonl
 """
 

examples/vlm/run_crafter_vlm_benchmark.py

@@ -224,7 +224,7 @@ async def _run_episode(
                         if unlocked:
                             achievements.add(str(name))
                 reward = obs.get("reward_last_step")
-                if isinstance(reward, (int, float)):
+                if isinstance(reward, int | float):
                     total_reward += float(reward)
 
             _save_observation_frame(env_response, frames_dir / f"step_{step_idx + 1:03d}.png")
@@ -263,7 +263,7 @@ def _summarise(results: list[EpisodeResult]) -> dict[str, Any]:
             "mean_steps": round(mean_steps, 2),
             "mean_achievements": round(mean_achievements, 2),
             "total_tool_calls": sum(r.tool_calls for r in mode_results),
-            "achievements": {name: count for name, count in sorted(achievement_counts.items())},
+            "achievements": dict(sorted(achievement_counts.items())),
         }
     return summary
 

examples/warming_up_to_rl/_utils.py

@@ -0,0 +1,92 @@
+from __future__ import annotations
+
+from collections.abc import Iterable, Sequence
+
+from synth_ai.task import (
+    RolloutEnvSpec,
+    RolloutPolicySpec,
+    RolloutRecordConfig,
+    RolloutRequest,
+    RolloutSafetyConfig,
+)
+
+DEFAULT_POLICY_NAME = "crafter-react"
+DEFAULT_ENV_NAME = "crafter"
+
+
+def parse_ops(spec: str | None) -> list[str] | None:
+    """Parse a comma-separated operations string into a list."""
+
+    if spec is None:
+        return None
+    ops = [op.strip() for op in spec.split(",") if op.strip()]
+    if not ops:
+        raise ValueError("Ops must contain at least one entry")
+    return ops
+
+
+def ops_from_pairs(max_llm_calls: int, *, cap: int | None = None) -> list[str]:
+    """Return alternating agent/env ops for the requested number of LLM calls."""
+
+    pairs = max(1, int(max_llm_calls or 0))
+    if cap is not None:
+        pairs = min(pairs, cap)
+    ops: list[str] = []
+    for _ in range(pairs):
+        ops.extend(["agent", "env"])
+    return ops
+
+
+def build_rollout_request(
+    *,
+    seed: int,
+    run_id: str,
+    model: str,
+    inference_url: str,
+    ops: Sequence[str] | Iterable[str],
+    inference_api_key: str | None = None,
+    extra_headers: dict[str, str] | None = None,
+    trace_format: str = "compact",
+    return_trace: bool = False,
+    policy_name: str = DEFAULT_POLICY_NAME,
+    env_name: str = DEFAULT_ENV_NAME,
+    max_policy_tokens: int | None = None,
+    record_trajectories: bool = True,
+) -> RolloutRequest:
+    """Construct a RolloutRequest shared across local rollout utilities."""
+
+    policy_config: dict[str, object] = {
+        "model": model,
+        "inference_url": inference_url,
+    }
+    if inference_api_key is not None:
+        policy_config["api_key"] = inference_api_key
+    if extra_headers:
+        policy_config["extra_headers"] = extra_headers
+    if max_policy_tokens is not None:
+        policy_config["max_completion_tokens"] = max_policy_tokens
+        policy_config["max_tokens"] = max_policy_tokens
+
+    record_cfg = RolloutRecordConfig(
+        trajectories=record_trajectories,
+        trace_format=trace_format,
+        return_trace=return_trace,
+    )
+    return RolloutRequest(
+        run_id=run_id,
+        env=RolloutEnvSpec(env_name=env_name, seed=seed, config={}),
+        policy=RolloutPolicySpec(policy_name=policy_name, config=policy_config),
+        ops=list(ops),
+        record=record_cfg,
+        on_done="reset",
+        safety=RolloutSafetyConfig(),
+    )
+
+
+__all__ = [
+    "DEFAULT_POLICY_NAME",
+    "DEFAULT_ENV_NAME",
+    "build_rollout_request",
+    "ops_from_pairs",
+    "parse_ops",
+]

examples/warming_up_to_rl/analyze_trace_db.py

@@ -383,7 +383,7 @@ def main() -> None:
     parser.add_argument(
         "--db",
         type=Path,
-        default=Path("traces/v3/synth_ai.db"),
+        default=Path("traces/task_app_traces.db"),
         help="Path to the tracing_v3 SQLite database",
     )
     args = parser.parse_args()

examples/warming_up_to_rl/configs/crafter_fft.toml

@@ -1,6 +1,8 @@
 # Crafter Full Finetune (FFT) example on H100
 # Adjust paths and hyperparameters to your environment before running.
 
+type = "sft"
+
 [job]
 model = "Qwen/Qwen3-4B"               # base model to finetune
 # Path to your SFT JSONL dataset

examples/warming_up_to_rl/configs/crafter_fft_4b.toml

@@ -1,5 +1,7 @@
 # FFT job config for Qwen/Qwen3-4B on Crafter SFT dataset
 
+type = "sft"
+
 [algorithm]
 type = "offline"
 method = "supervised_finetune"

examples/warming_up_to_rl/configs/eval_fft_qwen4b.toml

@@ -1,5 +1,7 @@
 # Eval config for finetuned Qwen/Qwen3-4B (FFT) via task app rollout
 
+type = "sft"
+
 # Required
 task_app_url = "https://synth-laboratories--grpo-crafter-task-app-final-warming--ceb5b2.modal.run"
 # Replace with your finished job id if different

examples/warming_up_to_rl/configs/eval_groq_qwen32b.toml

@@ -1,6 +1,8 @@
 # Eval config for Groq Qwen3-32B
 # Fields mirror run_eval.py expectations
 
+type = "rl"
+
 # Required
 task_app_url = "https://synth-laboratories--grpo-crafter-task-app-final-warming--ceb5b2.modal.run"
 model = "qwen/qwen3-32b"

examples/warming_up_to_rl/configs/eval_modal_qwen4b.toml

@@ -1,5 +1,7 @@
 # Eval config for Synth Modal inference Qwen/Qwen3-4B via task app rollout
 
+type = "rl"
+
 # Required
 task_app_url = "https://synth-laboratories--grpo-crafter-task-app-final-warming--ceb5b2.modal.run"
 model = "Qwen/Qwen3-4B"
@@ -20,4 +22,3 @@ concurrency = 10
 #   fetch the vLLM base from the task app /info to use as inference_url.
 # - Ensure the task app mounts the openai-api-key secret if your vLLM gateway
 #   requires a bearer token (OPENAI_API_KEY). Otherwise it will call unauthenticated.
-

examples/warming_up_to_rl/configs/rl_from_base_qwen4b.toml

@@ -1,11 +1,12 @@
 # RL training starting from base Qwen/Qwen3-4B (TOML-only model selection)
 
+type = "rl"
+
 [algorithm]
 type = "online"
 method = "policy_gradient"
 variety = "gspo"
 
-
 [services]
 task_url = "https://synth-laboratories--grpo-crafter-task-app-final-warming--ceb5b2.modal.run"
 

examples/warming_up_to_rl/configs/rl_from_ft.toml

@@ -1,5 +1,7 @@
 # RL training starting from a finetuned model id (TOML-only model selection)
 
+type = "rl"
+
 [services]
 # Task app base URL used by the RL job for rollouts
 # task_url = "https://YOUR-TASK-APP.modal.run"