hud-python 0.4.16__tar.gz → 0.4.17__tar.gz

{hud_python-0.4.16 → hud_python-0.4.17}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hud-python
-Version: 0.4.16
+Version: 0.4.17
 Summary: SDK for the HUD platform.
 Project-URL: Homepage, https://github.com/hud-evals/hud-python
 Project-URL: Bug Tracker, https://github.com/hud-evals/hud-python/issues

{hud_python-0.4.16 → hud_python-0.4.17}/hud/agents/claude.py

@@ -85,8 +85,8 @@ class ClaudeAgent(MCPAgent):
         self._claude_to_mcp_tool_map: dict[str, str] = {}
         self.claude_tools: list[dict] = []
 
-        # Base system prompt for autonomous operation
-        self.system_prompt = """
+        # Append Claude-specific instructions to the base system prompt
+        claude_instructions = """
         You are Claude, an AI assistant created by Anthropic. You are helpful, harmless, and honest.
         
         When working on tasks:
@@ -99,6 +99,12 @@ class ClaudeAgent(MCPAgent):
         Remember: You are expected to complete tasks autonomously. The user trusts you to accomplish what they asked.
         """.strip()  # noqa: E501
 
+        # Append Claude instructions to any base system prompt
+        if self.system_prompt:
+            self.system_prompt = f"{self.system_prompt}\n\n{claude_instructions}"
+        else:
+            self.system_prompt = claude_instructions
+
     async def initialize(self, task: str | Task | None = None) -> None:
         """Initialize the agent and build tool mappings."""
         await super().initialize(task)

{hud_python-0.4.16 → hud_python-0.4.17}/hud/agents/openai.py

@@ -78,8 +78,8 @@ class OperatorAgent(MCPAgent):
 
         self.model_name = "openai-" + self.model
 
-        # Base system prompt for autonomous operation
-        self.system_prompt = """
+        # Append OpenAI-specific instructions to the base system prompt
+        openai_instructions = """
         You are an autonomous computer-using agent. Follow these guidelines:
 
         1. NEVER ask for confirmation. Complete all tasks autonomously.
@@ -93,6 +93,12 @@ class OperatorAgent(MCPAgent):
         Remember: You are expected to complete tasks autonomously. The user trusts you to do what they asked.
         """.strip()  # noqa: E501
 
+        # Append OpenAI instructions to any base system prompt
+        if self.system_prompt:
+            self.system_prompt = f"{self.system_prompt}\n\n{openai_instructions}"
+        else:
+            self.system_prompt = openai_instructions
+
     async def _run_context(self, context: list[types.ContentBlock], max_steps: int = 10) -> Trace:
         """
         Run the agent with the given prompt or task.

{hud_python-0.4.16 → hud_python-0.4.17}/hud/cli/rl/__init__.py

@@ -23,7 +23,10 @@ def rl_main(
     ctx: typer.Context,
     model: str = typer.Option("Qwen/Qwen2.5-3B-Instruct", "--model", "-m", help="Model to train"),
     dataset: str | None = typer.Option(
-        None, "--dataset", "-d", help="Override dataset from lock file"
+        None,
+        "--dataset",
+        "-d",
+        help="Dataset: JSON file path or HuggingFace name (auto-detects if not provided)",
     ),
     config: Path | None = typer.Option(None, "--config", "-c", help="Config YAML path"),  # noqa: B008
     gpus: str = typer.Option("2xA100", "--gpus", help="GPU configuration (e.g., 2xA100, 4xH100)"),
@@ -39,9 +42,15 @@ def rl_main(
     3. Push environment to registry if needed
     4. Start remote training on Prime Intellect
 
+    Dataset can be:
+    - A local JSON file with tasks (e.g., tasks.json)
+    - A HuggingFace dataset name (e.g., 'username/dataset-name')
+    - Auto-detected from current directory if not specified
+
     Examples:
-        hud rl                    # Interactive mode with prompts
+        hud rl                    # Interactive mode, auto-detect tasks.json
         hud rl --model gpt2       # Train with specific model
+        hud rl --dataset tasks.json  # Use local task file
         hud rl --gpus 4xH100      # Use different GPU configuration
         hud rl init my-env:latest # Generate config for environment
     """

{hud_python-0.4.16 → hud_python-0.4.17}/hud/cli/rl/pod.py

@@ -62,6 +62,7 @@ async def create_and_connect_prime_pod(
     image: str,
     team_id: str | None = None,
     dataset_size: int | None = None,
+    is_json_file: bool = False,
 ) -> None:
     """Create a Prime Intellect pod and connect to it for training."""
     design.section_title("🌐 Creating Prime Intellect Pod")
@@ -330,6 +331,7 @@ async def create_and_connect_prime_pod(
                     output_dir=output_dir,
                     image=image,
                     dataset_size=dataset_size,
+                    is_json_file=is_json_file,
                 )
             else:
                 # Manual fallback
@@ -457,6 +459,7 @@ async def run_prime_training(
     auto_create_pod: str | None = None,
     team_id: str | None = None,
     dataset_size: int | None = None,
+    is_json_file: bool = False,
 ) -> None:
     """Run training on Prime Intellect infrastructure."""
     # Check API key
@@ -488,4 +491,5 @@ async def run_prime_training(
         image=image,
         team_id=team_id,
         dataset_size=dataset_size,
+        is_json_file=is_json_file,
     )

{hud_python-0.4.16 → hud_python-0.4.17}/hud/cli/rl/ssh.py

@@ -101,6 +101,7 @@ async def connect_and_train(
     output_dir: Path,
     image: str,
     dataset_size: int | None = None,
+    is_json_file: bool = False,
 ) -> None:
     """Connect to the pod via SSH and run training commands."""
     design.section_title("🚀 Starting Remote Training")
@@ -175,6 +176,37 @@ async def connect_and_train(
             design.info("Make sure scp is installed and in your PATH")
         raise typer.Exit(1) from e
 
+    # If dataset is a JSON file, copy it too
+    remote_dataset = dataset  # Default to unchanged
+    if is_json_file:
+        design.info("Copying task file to pod...")
+        try:
+            # On Windows, we need to ensure proper path formatting
+            dataset_path = str(dataset).replace("\\", "/")
+            # Extract just the filename for the remote path
+            dataset_filename = os.path.basename(dataset)
+            remote_dataset = f"/root/{dataset_filename}"
+
+            scp_cmd = [
+                "scp",
+                "-i",
+                str(ssh_key_path),
+                "-P",
+                ssh_port,
+                "-o",
+                "StrictHostKeyChecking=no",
+                "-o",
+                "UserKnownHostsFile=/dev/null",
+                dataset_path,
+                f"{ssh_user_host}:{remote_dataset}",
+            ]
+            design.debug(f"Running: {' '.join(scp_cmd)}")
+            subprocess.run(scp_cmd, check=True)  # noqa: S603, ASYNC221
+            design.success(f"Task file copied to {remote_dataset}")
+        except subprocess.CalledProcessError as e:
+            design.error(f"Failed to copy task file: {e}")
+            raise typer.Exit(1) from e
+
     design.info("Setting up environment and starting training...")
     design.info("This will take a few minutes for initial setup, then training will begin.")
     design.info("")
@@ -196,7 +228,7 @@ async def connect_and_train(
         "# Load environment",
         "env = vf.load_environment(",
         '    env_id="hud-vf-gym",',
-        f'    taskset="{dataset}",',
+        f'    taskset="{remote_dataset}",',
         '    config_path="/root/config.yaml",',
         f"    num_tasks={dataset_size},",
         ")",
@@ -242,7 +274,7 @@ async def connect_and_train(
         "uv venv --python 3.12 && "
         "source .venv/bin/activate && "
         # Install packages
-        "prime env install hud/hud-vf-gym@0.1.0 && "
+        "prime env install hud/hud-vf-gym@0.1.1 && "
         "uv pip install 'verifiers[train]' && "
         "uv pip install flash-attn --no-build-isolation && "
         # Set environment variables

{hud_python-0.4.16 → hud_python-0.4.17}/hud/cli/rl/train.py

@@ -23,6 +23,40 @@ from .utils import (
 design = HUDDesign()
 
 
+def find_task_json_files() -> list[Path]:
+    """Find JSON files containing tasks in the current directory."""
+    json_files = []
+    patterns = [
+        "*task*.json",
+        "*eval*.json",
+        "*Task*.json",
+        "*Eval*.json",
+        "*TASK*.json",
+        "*EVAL*.json",
+        "tasks.json",  # Most common name
+    ]
+
+    # First check current directory
+    for pattern in patterns:
+        json_files.extend(Path(".").glob(pattern))
+
+    # If no files found, search one level deep
+    if not json_files:
+        for pattern in patterns:
+            json_files.extend(Path(".").glob(f"*/{pattern}"))
+
+    # Remove duplicates and sort, prioritizing "tasks.json"
+    json_files = sorted(set(json_files))
+
+    # Put tasks.json first if it exists
+    tasks_json = Path("tasks.json")
+    if tasks_json in json_files:
+        json_files.remove(tasks_json)
+        json_files.insert(0, tasks_json)
+
+    return json_files
+
+
 def train_command_wrapper(
     model: str,
     dataset: str | None,
@@ -128,45 +162,22 @@ def train_command_wrapper(
                     raise typer.Exit(1)
 
         if "dataset" in missing:
-            # Check if we have tasks.json
-            tasks_file = Path("tasks.json")
-            if tasks_file.exists():
-                create_dataset = design.select(
-                    "Found tasks.json. Would you like to upload it as a dataset?",
-                    ["Yes, upload to HuggingFace", "No, I'll handle it manually"],
+            if missing["dataset"] == "multiple_json":
+                # Multiple JSON files found, let user choose
+                json_files = find_task_json_files()
+                design.info("Multiple task files found:")
+                file_choice = design.select(
+                    "Select a task file to use:",
+                    choices=[str(f) for f in json_files],
+                )
+                dataset = file_choice
+                design.success(f"Selected: {dataset}")
+            elif missing["dataset"] == "none":
+                design.error("No dataset specified and no task JSON files found")
+                design.info("Please use --dataset or create a tasks.json file")
+                design.hint(
+                    "Example: hud hf --name my-org/my-tasks  # Generate tasks from HUD evaluation"
                 )
-
-                if create_dataset == "Yes, upload to HuggingFace":
-                    dataset_name = typer.prompt("Enter dataset name (e.g., username/dataset-name)")
-
-                    if not validate_dataset_name(dataset_name):
-                        design.error("Invalid dataset name format. Expected: username/dataset-name")
-                        raise typer.Exit(1)
-
-                    design.info(f"Running 'hud hf tasks.json --name {dataset_name}'...")
-                    design.info("")
-
-                    # Run hf command
-                    result = subprocess.run(  # noqa: S603
-                        ["hud", "hf", "tasks.json", "--name", dataset_name],  # noqa: S607
-                        capture_output=True,
-                        text=True,
-                    )
-
-                    if result.returncode == 0:
-                        design.success("Dataset uploaded successfully")
-                        dataset = dataset_name
-                    else:
-                        design.error("Failed to upload dataset")
-                        if result.stderr:
-                            design.error(result.stderr)
-                        raise typer.Exit(1)
-                else:
-                    design.info("Please specify a dataset with --dataset")
-                    raise typer.Exit(1)
-            else:
-                design.error("No dataset specified and no tasks.json found")
-                design.info("Use --dataset to specify a HuggingFace dataset")
                 raise typer.Exit(1)
 
     # Ask about pod creation for Prime training
@@ -247,9 +258,123 @@ async def train_command(
         design.hint("Run 'hud build' first or specify with 'hud rl init <image>'")
         raise typer.Exit(1)
 
-    # Validate dataset has sufficient tasks for training
+    # Handle dataset (JSON file or HuggingFace dataset)
     dataset_size = None
-    if dataset:
+    is_json_file = False
+
+    # Use dataset from command or look for JSON files
+    if not dataset:
+        # Check for JSON files if no dataset specified
+        json_files = find_task_json_files()
+        if json_files:
+            if len(json_files) == 1:
+                dataset = str(json_files[0])
+                design.info(f"Found task file: {dataset}")
+                is_json_file = True
+            else:
+                # This case should have been handled in train_command_wrapper
+                design.error("Multiple task files found but none selected")
+                raise typer.Exit(1)
+        else:
+            # Use dataset from lock file
+            dataset = get_primary_dataset()
+            if dataset:
+                design.info(f"Using dataset from lock file: {dataset}")
+
+    # Check if dataset is a file path
+    if dataset and Path(dataset).exists() and dataset.endswith(".json"):
+        is_json_file = True
+
+    # Validate dataset
+    if dataset and is_json_file:
+        # Load and validate JSON file
+        design.info(f"Validating task file: {dataset}")
+        try:
+            with open(dataset) as f:  # noqa: ASYNC230
+                tasks_data = json.load(f)
+
+            # Handle both single task and array of tasks
+            if isinstance(tasks_data, dict):
+                tasks = [tasks_data]
+            elif isinstance(tasks_data, list):
+                tasks = tasks_data
+            else:
+                design.error("Invalid tasks file format")
+                raise typer.Exit(1)
+
+            dataset_size = len(tasks)
+            if dataset_size < 4:
+                design.error(f"Task file has only {dataset_size} tasks")
+                design.info("RL training requires at least 4 tasks for proper batching")
+                design.hint("Consider adding more tasks to your JSON file")
+                raise typer.Exit(1)
+
+            design.success(f"✓ Task file has {dataset_size} tasks")
+
+            # Check and convert MCP configs to remote if needed
+            if tasks:
+                sample_task = tasks[0]
+                sample_mcp_config = sample_task.get("mcp_config", {})
+
+                # Check if using local MCP configs
+                config_type = "unknown"
+                for server_config in sample_mcp_config.values():
+                    if isinstance(server_config, dict) and "url" in server_config:
+                        url = server_config.get("url", "")
+                        if "mcp.hud.so" in url:
+                            config_type = "remote"
+                            break
+                        else:
+                            config_type = "local"
+
+                if config_type == "local":
+                    design.info("Converting local MCP configs to remote for training...")
+
+                    # Get the image name from lock file or environment
+                    from .utils import get_image_from_lock
+
+                    env_image = image or get_image_from_lock()
+
+                    if not env_image:
+                        design.error("No image found for remote MCP conversion")
+                        design.hint("Run 'hud build' first")
+                        raise typer.Exit(1)
+
+                    # Check if image needs to be pushed
+                    if "/" not in env_image or env_image.startswith("local/"):
+                        design.warning(f"Image '{env_image}' appears to be local only")
+                        design.info("Running 'hud push' to make it publicly available...")
+                        from hud.cli.push import push_command
+
+                        push_command(directory=".", yes=True)
+                        design.success("Image pushed successfully")
+                        # Re-read image name after push
+                        env_image = get_image_from_lock()
+
+                    # Convert all tasks to use remote MCP
+                    for task in tasks:
+                        remote_config = {
+                            "hud": {
+                                "url": "https://mcp.hud.so/v3/mcp",
+                                "headers": {
+                                    "Authorization": "Bearer $HUD_API_KEY",
+                                    "Mcp-Image": env_image,
+                                },
+                            }
+                        }
+                        task["mcp_config"] = remote_config
+
+                    design.success("✓ Converted all tasks to use remote MCP configs")
+
+                    # Save the modified tasks back to the file
+                    with open(dataset, "w") as f:  # noqa: ASYNC230
+                        json.dump(tasks, f, indent=2)
+                    design.info("Updated task file with remote configs")
+        except json.JSONDecodeError as e:
+            design.error(f"Invalid JSON in task file: {e}")
+            raise typer.Exit(1) from e
+    elif dataset:
+        # Validate HuggingFace dataset
         design.info(f"Validating dataset: {dataset}")
         try:
             # Try to load dataset info from HuggingFace
@@ -273,12 +398,6 @@ async def train_command(
             design.warning(f"Could not validate dataset size: {e}")
             design.info("Proceeding with training - ensure dataset has at least 4 tasks")
 
-    # Use dataset from command or lock file
-    if not dataset:
-        dataset = get_primary_dataset()
-        if dataset:
-            design.info(f"Using dataset from lock file: {dataset}")
-
     # Display configuration
     design.section_title("📋 Training Configuration")
     design.json_config(
@@ -318,6 +437,7 @@ async def train_command(
         auto_create_pod=auto_create_pod,
         team_id=team_id,
         dataset_size=dataset_size,
+        is_json_file=is_json_file,
     )
 
 
@@ -340,10 +460,19 @@ def check_requirements(config: Path | None, dataset: str | None) -> dict[str, An
 
     # Check dataset
     if not dataset:
-        # Check lock file for dataset
-        primary_dataset = get_primary_dataset()
-        if not primary_dataset:
-            missing["dataset"] = "none"
+        # First check for JSON files (preferred method)
+        json_files = find_task_json_files()
+        if json_files:
+            if len(json_files) == 1:
+                # Will be auto-selected
+                pass
+            else:
+                missing["dataset"] = "multiple_json"
+        else:
+            # Check lock file for HuggingFace dataset
+            primary_dataset = get_primary_dataset()
+            if not primary_dataset:
+                missing["dataset"] = "none"
 
     return missing
 
@@ -407,13 +536,23 @@ async def run_remote_training(
     auto_create_pod: str | None = None,
     team_id: str | None = None,
     dataset_size: int | None = None,
+    is_json_file: bool = False,
 ) -> None:
     """Run training on remote infrastructure."""
     design.section_title("🚀 Remote Training")
 
     if provider == "prime":
         await run_prime_training(
-            model, dataset, config, gpus, output_dir, image, auto_create_pod, team_id, dataset_size
+            model,
+            dataset,
+            config,
+            gpus,
+            output_dir,
+            image,
+            auto_create_pod,
+            team_id,
+            dataset_size,
+            is_json_file,
         )
     else:
         design.error(f"Provider '{provider}' not yet supported")

{hud_python-0.4.16 → hud_python-0.4.17}/hud/utils/tests/test_version.py

@@ -5,4 +5,4 @@ def test_import():
     """Test that the package can be imported."""
     import hud
 
-    assert hud.__version__ == "0.4.16"
+    assert hud.__version__ == "0.4.17"

{hud_python-0.4.16 → hud_python-0.4.17}/hud/version.py

@@ -4,4 +4,4 @@ Version information for the HUD SDK.
 
 from __future__ import annotations
 
-__version__ = "0.4.16"
+__version__ = "0.4.17"

{hud_python-0.4.16 → hud_python-0.4.17}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hud-python"
-version = "0.4.16"
+version = "0.4.17"
 description = "SDK for the HUD platform."
 readme = "README.md"
 requires-python = ">=3.11, <3.14"