inspect-ai 0.3.73__py3-none-any.whl → 0.3.75__py3-none-any.whl

inspect_ai/_eval/task/task.py

@@ -13,6 +13,7 @@ from inspect_ai.approval._policy import ApprovalPolicy, approval_policies_from_c
 from inspect_ai.dataset import Dataset, MemoryDataset, Sample
 from inspect_ai.log import EvalLog
 from inspect_ai.model import GenerateConfig
+from inspect_ai.model._model import Model, get_model
 from inspect_ai.scorer import Metric, Scorer
 from inspect_ai.scorer._reducer import ScoreReducers, create_reducers
 from inspect_ai.solver import Plan, Solver, generate
@@ -50,6 +51,7 @@ class Task:
         cleanup: Callable[[TaskState], Awaitable[None]] | None = None,
         scorer: Scorer | list[Scorer] | None = None,
         metrics: list[Metric] | dict[str, list[Metric]] | None = None,
+        model: str | Model | None = None,
         config: GenerateConfig = GenerateConfig(),
         sandbox: SandboxEnvironmentType | None = None,
         approval: str | list[ApprovalPolicy] | None = None,
@@ -67,42 +69,38 @@ class Task:
         """Create a task.
 
         Args:
-            dataset (Dataset | Sequence[Sample]): Dataset to evaluate
-            setup: (Solver | list[Solver] | None): Setup step (always run
-                even when the main `solver` is replaced).
-            solver: (Solver | list[Solver]): Solver or list of solvers.
-                Defaults to generate(), a normal call to the model.
+            dataset: Dataset to evaluate
+            setup: Setup step (always run even when the main `solver` is replaced).
+            solver: Solver or list of solvers. Defaults to generate(), a normal call to the model.
             cleanup: Optional cleanup function for task. Called after
                 all solvers have run for each sample (including if an
                 exception occurs during the run)
-            scorer: (Scorer | list[Scorer] | None): Scorer used to evaluate model output.
-            metrics (list[Metric] | dict[str, list[Metric]] | None):
-                Alternative metrics (overrides the metrics provided by the specified scorer).
-            config (GenerateConfig): Model generation config.
-            sandbox (SandboxEnvironmentType | None): Sandbox environment type
-                (or optionally a str or tuple with a shorthand spec)
-            approval: (str | list[ApprovalPolicy] | None): Tool use approval policies.
-                Either a path to an approval policy config file or a list of approval policies.
-                Defaults to no approval policy.
-            epochs (int | Epochs | None): Epochs to repeat samples for and optional score
+            scorer: Scorer used to evaluate model output.
+            metrics: Alternative metrics (overrides the metrics provided by the specified scorer).
+            model: Default model for task (Optional, defaults to eval model).
+            config: Model generation config.
+            sandbox: Sandbox environment type (or optionally a str or tuple with a shorthand spec)
+            approval: Tool use approval policies.
+                Either a path to an approval policy config file or a list of approval policies. Defaults to no approval policy.
+            epochs: Epochs to repeat samples for and optional score
                 reducer function(s) used to combine sample scores (defaults to "mean")
-            fail_on_error (bool | float | None): `True` to fail on first sample error
+            fail_on_error: `True` to fail on first sample error
                 (default); `False` to never fail on sample errors; Value between 0 and 1
                 to fail if a proportion of total samples fails. Value greater than 1 to fail
                 eval if a count of samples fails.
-            message_limit (int | None): Limit on total messages used for each sample.
-            token_limit (int | None): Limit on total tokens used for each sample.
+            message_limit: Limit on total messages used for each sample.
+            token_limit: Limit on total tokens used for each sample.
             time_limit: Limit on clock time (in seconds) for samples.
             working_limit: Limit on working time (in seconds) for sample. Working
                 time includes model generation, tool calls, etc. but does not include
                 time spent waiting on retries or shared resources.
-            name: (str | None): Task name. If not specified is automatically
+            name: Task name. If not specified is automatically
                 determined based on the name of the task directory (or "task")
                 if its anonymous task (e.g. created in a notebook and passed to
                 eval() directly)
-            version: (int): Version of task (to distinguish evolutions
+            version: Version of task (to distinguish evolutions
                 of the task spec or breaking changes to it)
-            metadata: (dict[str, Any] | None): Additional metadata to associate with the task.
+            metadata:  Additional metadata to associate with the task.
             **kwargs: Deprecated arguments.
         """
         # handle deprecated args
@@ -135,6 +133,7 @@ class Task:
         self.cleanup = cleanup
         self.scorer = resolve_scorer(scorer)
         self.metrics = metrics
+        self.model = resolve_model(model)
         self.config = config
         self.sandbox = resolve_sandbox_environment(sandbox)
         self.approval = resolve_approval(approval)
@@ -176,6 +175,7 @@ def task_with(
     cleanup: Callable[[TaskState], Awaitable[None]] | None | NotGiven = NOT_GIVEN,
     scorer: Scorer | list[Scorer] | None | NotGiven = NOT_GIVEN,
     metrics: list[Metric] | dict[str, list[Metric]] | None | NotGiven = NOT_GIVEN,
+    model: str | Model | NotGiven = NOT_GIVEN,
     config: GenerateConfig | NotGiven = NOT_GIVEN,
     sandbox: SandboxEnvironmentType | None | NotGiven = NOT_GIVEN,
     approval: str | list[ApprovalPolicy] | None | NotGiven = NOT_GIVEN,
@@ -192,43 +192,39 @@ def task_with(
     """Task adapted with alternate values for one or more options.
 
     Args:
-        task (Task): Task to adapt (it is deep copied prior to mutating options)
-        dataset (Dataset | Sequence[Sample]): Dataset to evaluate
-        setup: (Solver | list[Solver] | None): Setup step (always run
-            even when the main `solver` is replaced).
-        solver: (Solver | list[Solver]): Solver or list of solvers.
-            Defaults to generate(), a normal call to the model.
+        task: Task to adapt (it is deep copied prior to mutating options)
+        dataset: Dataset to evaluate
+        setup: Setup step (always run even when the main `solver` is replaced).
+        solver: Solver or list of solvers. Defaults to generate(), a normal call to the model.
         cleanup: Optional cleanup function for task. Called after
             all solvers have run for each sample (including if an
             exception occurs during the run)
-        scorer: (Scorer | list[Scorer] | None): Scorer used to evaluate model output.
-        metrics (list[Metric] | dict[str, list[Metric]] | None):
-            Alternative metrics (overrides the metrics provided by the specified scorer).
-        config (GenerateConfig): Model generation config.
-        sandbox (SandboxEnvironmentType | None): Sandbox environment type
-            (or optionally a str or tuple with a shorthand spec)
-        approval: (str | list[ApprovalPolicy] | None): Tool use approval policies.
-            Either a path to an approval policy config file or a list of approval policies.
-            Defaults to no approval policy.
-        epochs (int | Epochs | None): Epochs to repeat samples for and optional score
+        scorer: Scorer used to evaluate model output.
+        metrics: Alternative metrics (overrides the metrics provided by the specified scorer).
+        model: Default model for task (Optional, defaults to eval model).
+        config: Model generation config.
+        sandbox: Sandbox environment type (or optionally a str or tuple with a shorthand spec)
+        approval: Tool use approval policies.
+            Either a path to an approval policy config file or a list of approval policies. Defaults to no approval policy.
+        epochs: Epochs to repeat samples for and optional score
             reducer function(s) used to combine sample scores (defaults to "mean")
-        fail_on_error (bool | float | None): `True` to fail on first sample error
+        fail_on_error: `True` to fail on first sample error
             (default); `False` to never fail on sample errors; Value between 0 and 1
             to fail if a proportion of total samples fails. Value greater than 1 to fail
             eval if a count of samples fails.
-        message_limit (int | None): Limit on total messages used for each sample.
-        token_limit (int | None): Limit on total tokens used for each sample.
+        message_limit: Limit on total messages used for each sample.
+        token_limit: Limit on total tokens used for each sample.
         time_limit: Limit on clock time (in seconds) for samples.
-        working_limit: Limit on execution time (in seconds) for sample. Execution
+        working_limit: Limit on working time (in seconds) for sample. Working
             time includes model generation, tool calls, etc. but does not include
             time spent waiting on retries or shared resources.
-        name: (str | None): Task name. If not specified is automatically
+        name: Task name. If not specified is automatically
             determined based on the name of the task directory (or "task")
             if its anonymous task (e.g. created in a notebook and passed to
             eval() directly)
-        version: (int): Version of task (to distinguish evolutions
+        version: Version of task (to distinguish evolutions
             of the task spec or breaking changes to it)
-        metadata: (dict[str, Any] | None): Additional metadata to associate with the task.
+        metadata:  Additional metadata to associate with the task.
 
     Returns:
         Task: Task adapted with alternate options.
@@ -248,6 +244,8 @@ def task_with(
         task.scorer = resolve_scorer(scorer)
     if not isinstance(metrics, NotGiven):
         task.metrics = metrics
+    if not isinstance(model, NotGiven):
+        task.model = resolve_model(model)
     if not isinstance(config, NotGiven):
         task.config = config
     if not isinstance(sandbox, NotGiven):
@@ -307,34 +305,10 @@ class PreviousTask:
     id: str
     task: str | Task
     task_args: dict[str, Any]
+    model: Model | None
     log: EvalLog
 
 
-Tasks = (
-    str
-    | PreviousTask
-    | TaskInfo
-    | Task
-    | Callable[..., Task]
-    | type[Task]
-    | list[str]
-    | list[PreviousTask]
-    | list[TaskInfo]
-    | list[Task]
-    | list[Callable[..., Task]]
-    | list[type[Task]]
-    | None
-)
-r"""One or more tasks.
-
-Tasks to be evaluated. Many forms of task specification are
-supported including directory names, task functions, task
-classes, and task instances (a single task or list of tasks
-can be specified). None is a request to read a task out
-of the current working directory.
-"""
-
-
 def resolve_approval(
     approval: str | list[ApprovalPolicy] | None,
 ) -> list[ApprovalPolicy] | None:
@@ -370,6 +344,13 @@ def resolve_solver(solver: Solver | list[Solver]) -> Solver:
     return chain(solver) if isinstance(solver, list) else solver
 
 
+def resolve_model(model: str | Model | None) -> Model | None:
+    if isinstance(model, str):
+        return get_model(model)
+    else:
+        return model
+
+
 def resolve_scorer(scorer: Scorer | list[Scorer] | None) -> list[Scorer] | None:
     return (
         scorer if isinstance(scorer, list) else [scorer] if scorer is not None else None

inspect_ai/_eval/task/tasks.py

@@ -0,0 +1,30 @@
+from typing import Callable, TypeAlias
+
+from .resolved import ResolvedTask
+from .task import PreviousTask, Task, TaskInfo
+
+Tasks: TypeAlias = (
+    str
+    | PreviousTask
+    | ResolvedTask
+    | TaskInfo
+    | Task
+    | Callable[..., Task]
+    | type[Task]
+    | list[str]
+    | list[PreviousTask]
+    | list[ResolvedTask]
+    | list[TaskInfo]
+    | list[Task]
+    | list[Callable[..., Task]]
+    | list[type[Task]]
+    | None
+)
+r"""One or more tasks.
+
+Tasks to be evaluated. Many forms of task specification are
+supported including directory names, task functions, task
+classes, and task instances (a single task or list of tasks
+can be specified). None is a request to read a task out
+of the current working directory.
+"""

inspect_ai/_util/constants.py

@@ -36,3 +36,6 @@ CONSOLE_DISPLAY_WIDTH = 120
 BASE_64_DATA_REMOVED = "<base64-data-removed>"
 SANDBOX_SETUP_TIMEOUT = 300
 NO_CONTENT = "(no content)"
+
+DESERIALIZING = "deserializing"
+DESERIALIZING_CONTEXT = {DESERIALIZING: True}

inspect_ai/_util/dotenv.py

@@ -52,6 +52,9 @@ def init_dotenv() -> None:
         if inspect_log_dir:
             os.environ[INSPECT_LOG_DIR_VAR] = inspect_log_dir
 
+        # re-apply any env vars specified at the cli w/ --env
+        apply_cli_env()
+
 
 @contextlib.contextmanager
 def dotenv_environ(
@@ -76,3 +79,17 @@ def dotenv_environ(
     finally:
         os.environ.update(update_after)
         [os.environ.pop(k) for k in remove_after]
+
+
+_cli_env: dict[str, Any] = {}
+
+
+def init_cli_env(env: dict[str, Any]) -> None:
+    global _cli_env
+    _cli_env = env
+    apply_cli_env()
+
+
+def apply_cli_env() -> None:
+    for var, value in _cli_env.items():
+        os.environ[var] = str(value)

inspect_ai/_util/registry.py

@@ -5,6 +5,7 @@ from typing import Any, Callable, Literal, TypedDict, TypeGuard, cast
 from pydantic import BaseModel, Field
 from pydantic_core import to_jsonable_python
 
+from inspect_ai._util.json import jsonable_python
 from inspect_ai._util.package import get_installed_package_name
 
 from .constants import PKG_NAME
@@ -198,13 +199,15 @@ def registry_create(type: RegistryType, name: str, **kwargs: Any) -> object:
     def with_registry_info(o: object) -> object:
         return set_registry_info(o, registry_info(obj))
 
-    # instantiate registry objects
+    # instantiate registry and model objects
     for param in kwargs.keys():
         value = kwargs[param]
         if is_registry_dict(value):
             kwargs[param] = registry_create(
                 value["type"], value["name"], **value["params"]
             )
+        elif is_model_dict(value):
+            kwargs[param] = model_create_from_dict(value)
 
     if isclass(obj):
         return with_registry_info(obj(**kwargs))
@@ -380,6 +383,8 @@ def is_registry_dict(o: object) -> TypeGuard[RegistryDict]:
 
 
 def registry_value(o: object) -> Any:
+    from inspect_ai.model._model import Model
+
     # treat tuple as list
     if isinstance(o, tuple):
         o = list(o)
@@ -390,14 +395,50 @@ def registry_value(o: object) -> Any:
     elif isinstance(o, dict):
         return {k: registry_value(v) for k, v in o.items()}
     elif has_registry_params(o):
-        return dict(
+        return RegistryDict(
             type=registry_info(o).type,
             name=registry_log_name(o),
             params=registry_params(o),
         )
+    elif isinstance(o, Model):
+        return ModelDict(
+            model=str(o),
+            config=jsonable_python(o.config),
+            base_url=o.api.base_url,
+            model_args=o.model_args,
+        )
     else:
         return o
 
 
 def registry_create_from_dict(d: RegistryDict) -> object:
     return registry_create(d["type"], d["name"], **d["params"])
+
+
+class ModelDict(TypedDict):
+    model: str
+    config: dict[str, Any]
+    base_url: str | None
+    model_args: dict[str, Any]
+
+
+def is_model_dict(o: object) -> TypeGuard[ModelDict]:
+    return (
+        isinstance(o, dict)
+        and "model" in o
+        and "config" in o
+        and "base_url" in o
+        and "model_args" in o
+    )
+
+
+def model_create_from_dict(d: ModelDict) -> object:
+    from inspect_ai.model._generate_config import GenerateConfig
+    from inspect_ai.model._model import get_model
+
+    return get_model(
+        d["model"],
+        config=GenerateConfig(**d["config"]),
+        base_url=d["base_url"],
+        **d["model_args"],
+    )

inspect_ai/_view/server.py

@@ -57,8 +57,7 @@ def view_server(
     @routes.get("/api/logs/{log}")
     async def api_log(request: web.Request) -> web.Response:
         # log file requested
-        file = request.match_info["log"]
-        file = urllib.parse.unquote(file)
+        file = normalize_uri(request.match_info["log"])
         validate_log_file_request(file)
 
         # header_only is based on a size threshold
@@ -68,8 +67,7 @@ def view_server(
     @routes.get("/api/log-size/{log}")
     async def api_log_size(request: web.Request) -> web.Response:
         # log file requested
-        file = request.match_info["log"]
-        file = urllib.parse.unquote(file)
+        file = normalize_uri(request.match_info["log"])
         validate_log_file_request(file)
 
         return await log_size_response(file)
@@ -77,8 +75,7 @@ def view_server(
     @routes.get("/api/log-delete/{log}")
     async def api_log_delete(request: web.Request) -> web.Response:
         # log file requested
-        file = request.match_info["log"]
-        file = urllib.parse.unquote(file)
+        file = normalize_uri(request.match_info["log"])
         validate_log_file_request(file)
 
         return await log_delete_response(file)
@@ -86,8 +83,7 @@ def view_server(
     @routes.get("/api/log-bytes/{log}")
     async def api_log_bytes(request: web.Request) -> web.Response:
         # log file requested
-        file = request.match_info["log"]
-        file = urllib.parse.unquote(file)
+        file = normalize_uri(request.match_info["log"])
         validate_log_file_request(file)
 
         # header_only is based on a size threshold
@@ -106,7 +102,7 @@ def view_server(
         if authorization:
             request_log_dir = request.query.getone("log_dir", None)
             if request_log_dir:
-                request_log_dir = urllib.parse.unquote(request_log_dir)
+                request_log_dir = normalize_uri(request_log_dir)
             else:
                 request_log_dir = log_dir
         else:
@@ -121,7 +117,7 @@ def view_server(
     @routes.get("/api/log-headers")
     async def api_log_headers(request: web.Request) -> web.Response:
         files = request.query.getall("file", [])
-        files = [urllib.parse.unquote(file) for file in files]
+        files = [normalize_uri(file) for file in files]
         map(validate_log_file_request, files)
         return await log_headers_response(files)
 
@@ -166,6 +162,28 @@ def view_server(
     )
 
 
+def normalize_uri(uri: str) -> str:
+    """Normalize incoming URIs to a consistent format."""
+    # Decode any URL-encoded characters
+    parsed = urllib.parse.urlparse(urllib.parse.unquote(uri))
+
+    if parsed.scheme != "file":
+        # If this isn't a file uri, just unquote it
+        return urllib.parse.unquote(uri)
+
+    else:
+        # If this is a file uri, see whether we should process triple slashes
+        # down to double slashes
+        path = parsed.path
+
+        # Detect and normalize Windows-style file URIs
+        if path.startswith("/") and len(path) > 3 and path[2] == ":":
+            # Strip leading `/` before drive letter
+            path = path[1:]
+
+        return f"file://{path}"
+
+
 def log_listing_response(logs: list[EvalLogInfo], log_dir: str) -> web.Response:
     response = dict(
         log_dir=aliased_path(log_dir),

inspect_ai/_view/www/dist/assets/index.css

@@ -16346,7 +16346,7 @@ ul.jsondiffpatch-textdiff {
   column-gap: 0.5em;
   min-width: 200px;
 }
-._flatBody_gk2ju_1 {
+._flatBody_1uw6w_1 {
   color: var(--bs-danger);
   display: grid;
   grid-template-columns: max-content max-content;
@@ -16354,16 +16354,17 @@ ul.jsondiffpatch-textdiff {
   margin-top: 0.4rem;
 }
 
-._iconSmall_gk2ju_9 {
+._iconSmall_1uw6w_9 {
   font-size: var(--inspect-font-size-small);
   line-height: var(--inspect-font-size-small);
   height: var(--inspect-font-size-small);
 }
 
-._lineBase_gk2ju_15 {
+._lineBase_1uw6w_15 {
   font-size: var(--inspect-font-size-base);
   line-height: var(--inspect-font-size-base);
   height: var(--inspect-font-size-base);
+  max-width: 30em;
 }
 ._target_9qy4e_1 {
   padding-left: 0;