inspect-ai 0.3.76__py3-none-any.whl → 0.3.77__py3-none-any.whl

inspect_ai/_cli/eval.py

@@ -115,6 +115,13 @@ def eval_options(func: Callable[..., Any]) -> Callable[..., click.Context]:
         help="Tags to associate with this evaluation run.",
         envvar="INSPECT_EVAL_TAGS",
     )
+    @click.option(
+        "--metadata",
+        multiple=True,
+        type=str,
+        help="Metadata to associate with this evaluation run (more than one --metadata argument can be specified).",
+        envvar="INSPECT_EVAL_METADATA",
+    )
     @click.option(
         "--trace",
         type=bool,
@@ -449,6 +456,7 @@ def eval_command(
     s: tuple[str] | None,
     solver_config: str | None,
     tags: str | None,
+    metadata: tuple[str] | None,
     trace: bool | None,
     approval: str | None,
     sandbox: str | None,
@@ -525,6 +533,7 @@ def eval_command(
         s=s,
         solver_config=solver_config,
         tags=tags,
+        metadata=metadata,
         trace=trace,
         approval=approval,
         sandbox=sandbox,
@@ -616,6 +625,7 @@ def eval_set_command(
     s: tuple[str] | None,
     solver_config: str | None,
     tags: str | None,
+    metadata: tuple[str] | None,
     sandbox: str | None,
     no_sandbox_cleanup: bool | None,
     epochs: int | None,
@@ -695,6 +705,7 @@ def eval_set_command(
         s=s,
         solver_config=solver_config,
         tags=tags,
+        metadata=metadata,
         trace=trace,
         approval=approval,
         sandbox=sandbox,
@@ -749,6 +760,7 @@ def eval_exec(
     s: tuple[str] | None,
     solver_config: str | None,
     tags: str | None,
+    metadata: tuple[str] | None,
     trace: bool | None,
     approval: str | None,
     sandbox: str | None,
@@ -790,6 +802,9 @@ def eval_exec(
     # parse tags
     eval_tags = parse_comma_separated(tags)
 
+    # parse metadata
+    eval_metadata = parse_cli_args(metadata)
+
     # resolve epochs
     eval_epochs = (
         Epochs(epochs, create_reducers(parse_comma_separated(epochs_reducer)))
@@ -825,6 +840,7 @@ def eval_exec(
             task_args=task_args,
             solver=SolverSpec(solver, solver_args) if solver else None,
             tags=eval_tags,
+            metadata=eval_metadata,
             trace=trace,
             approval=approval,
             sandbox=parse_sandbox(sandbox),

inspect_ai/_display/core/results.py

@@ -131,9 +131,14 @@ def task_stats(stats: EvalStats) -> RenderableType:
         else:
             input_tokens = f"[bold]I: [/bold]{usage.input_tokens:,}"
 
+        if usage.reasoning_tokens is not None:
+            reasoning_tokens = f", [bold]R: [/bold]{usage.reasoning_tokens:,}"
+        else:
+            reasoning_tokens = ""
+
         table.add_row(
             Text(model, style="bold"),
-            f"  {usage.total_tokens:,} tokens [{input_tokens}, [bold]O: [/bold]{usage.output_tokens:,}]",
+            f"  {usage.total_tokens:,} tokens [{input_tokens}, [bold]O: [/bold]{usage.output_tokens:,}{reasoning_tokens}]",
             style=theme.light,
         )
 

inspect_ai/_eval/eval.py

@@ -68,6 +68,7 @@ def eval(
     sandbox_cleanup: bool | None = None,
     solver: Solver | list[Solver] | SolverSpec | None = None,
     tags: list[str] | None = None,
+    metadata: dict[str, Any] | None = None,
     trace: bool | None = None,
     display: DisplayType | None = None,
     approval: str | list[ApprovalPolicy] | None = None,
@@ -116,6 +117,7 @@ def eval(
         solver: Alternative solver for task(s).
             Optional (uses task solver by default).
         tags: Tags to associate with this evaluation run.
+        metadata: Metadata to associate with this evaluation run.
         trace: Trace message interactions with evaluated model to terminal.
         display: Task display type (defaults to 'full').
         approval: Tool use approval policies.
@@ -186,6 +188,7 @@ def eval(
                 sandbox_cleanup=sandbox_cleanup,
                 solver=solver,
                 tags=tags,
+                metadata=metadata,
                 approval=approval,
                 log_level=log_level,
                 log_level_transcript=log_level_transcript,
@@ -235,6 +238,7 @@ async def eval_async(
     sandbox_cleanup: bool | None = None,
     solver: Solver | list[Solver] | SolverSpec | None = None,
     tags: list[str] | None = None,
+    metadata: dict[str, Any] | None = None,
     approval: str | list[ApprovalPolicy] | ApprovalPolicyConfig | None = None,
     log_level: str | None = None,
     log_level_transcript: str | None = None,
@@ -274,7 +278,8 @@ async def eval_async(
         sandbox: Sandbox environment type (or optionally a str or tuple with a shorthand spec)
         sandbox_cleanup: Cleanup sandbox environments after task completes (defaults to True)
         solver: Alternative solver for task(s).  Optional (uses task solver by default).
-        tags (list[str] | None): Tags to associate with this evaluation run.
+        tags: Tags to associate with this evaluation run.
+        metadata: Metadata to associate with this evaluation run.
         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.
@@ -449,6 +454,7 @@ async def eval_async(
                         epochs_reducer=epochs_reducer,
                         solver=solver,
                         tags=tags,
+                        metadata=metadata,
                         score=score,
                         debug_errors=debug_errors is True,
                         **kwargs,
@@ -473,6 +479,7 @@ async def eval_async(
                 epochs_reducer=epochs_reducer,
                 solver=solver,
                 tags=tags,
+                metadata=metadata,
                 score=score,
                 **kwargs,
             )

inspect_ai/_eval/evalset.py

@@ -68,6 +68,7 @@ def eval_set(
     sandbox_cleanup: bool | None = None,
     solver: Solver | list[Solver] | SolverSpec | None = None,
     tags: list[str] | None = None,
+    metadata: dict[str, Any] | None = None,
     trace: bool | None = None,
     display: DisplayType | None = None,
     approval: str | list[ApprovalPolicy] | None = None,
@@ -127,6 +128,7 @@ def eval_set(
         solver: Alternative solver(s) for
             evaluating task(s). ptional (uses task solver by default).
         tags: Tags to associate with this evaluation run.
+        metadata: Metadata to associate with this evaluation run.
         trace: Trace message interactions with evaluated model to terminal.
         display: Task display type (defaults to 'full').
         approval: Tool use approval policies.
@@ -193,6 +195,7 @@ def eval_set(
             sandbox_cleanup=sandbox_cleanup,
             solver=solver,
             tags=tags,
+            metadata=metadata,
             trace=trace,
             display=display,
             approval=approval,

inspect_ai/_eval/run.py

@@ -2,7 +2,7 @@ import functools
 import logging
 import os
 import sys
-from typing import Awaitable, Callable, Set, cast
+from typing import Any, Awaitable, Callable, Set, cast
 
 from inspect_ai._eval.task.task import Task
 from inspect_ai._util.trace import trace_action
@@ -68,6 +68,7 @@ async def eval_run(
     epochs_reducer: list[ScoreReducer] | None = None,
     solver: Solver | SolverSpec | None = None,
     tags: list[str] | None = None,
+    metadata: dict[str, Any] | None = None,
     debug_errors: bool = False,
     score: bool = True,
     **kwargs: Unpack[GenerateConfigArgs],
@@ -205,7 +206,7 @@ async def eval_run(
                     task_args=resolved_task.task_args,
                     model_args=resolved_task.model.model_args,
                     eval_config=task_eval_config,
-                    metadata=task.metadata,
+                    metadata=((metadata or {}) | (task.metadata or {})) or None,
                     recorder=recorder,
                 )
                 await logger.init()

inspect_ai/_util/content.py

@@ -12,6 +12,9 @@ class ContentText(BaseModel):
     text: str
     """Text content."""
 
+    refusal: bool | None = Field(default=None)
+    """Was this a refusal message?"""
+
 
 class ContentReasoning(BaseModel):
     """Reasoning content.

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

@@ -21577,7 +21577,7 @@ var require_assets = __commonJS({
           className: clsx(
             "source-code",
             "sourceCode",
-            `language-${highlightLanguage}`,
+            highlightLanguage ? `language-${highlightLanguage}` : void 0,
             styles$10.outputCode
           ),
           children: formattedContent
@@ -21613,6 +21613,22 @@ var require_assets = __commonJS({
       }
       const collapse = Array.isArray(output2) ? output2.every((item2) => !isContentImage(item2)) : !isContentImage(output2);
       const normalizedContent = reactExports.useMemo(() => normalizeContent$1(output2), [output2]);
+      const hasContent = normalizedContent.find((c2) => {
+        if (c2.type === "tool") {
+          for (const t2 of c2.content) {
+            if (t2.type === "text") {
+              if (t2.text) {
+                return true;
+              }
+            } else {
+              return true;
+            }
+          }
+          return false;
+        } else {
+          return true;
+        }
+      });
       const contents2 = mode !== "compact" ? input2 : input2 || functionCall;
       return /* @__PURE__ */ jsxRuntimeExports.jsxs("div", { children: [
         mode !== "compact" && (!view || view.title) ? /* @__PURE__ */ jsxRuntimeExports.jsx(ToolTitle, { title: (view == null ? void 0 : view.title) || functionCall }) : "",
@@ -21625,7 +21641,7 @@ var require_assets = __commonJS({
               toolCallView: view
             }
           ),
-          /* @__PURE__ */ jsxRuntimeExports.jsx(ExpandablePanel, { collapse, border: true, lines: 15, children: /* @__PURE__ */ jsxRuntimeExports.jsx(MessageContent, { contents: normalizedContent }) })
+          hasContent ? /* @__PURE__ */ jsxRuntimeExports.jsx(ExpandablePanel, { collapse, border: true, lines: 15, children: /* @__PURE__ */ jsxRuntimeExports.jsx(MessageContent, { contents: normalizedContent }) }) : void 0
         ] }) })
       ] });
     };

inspect_ai/_view/www/src/samples/chat/tools/ToolCallView.tsx

@@ -83,8 +83,24 @@ export const ToolCallView: FC<ToolCallViewProps> = ({
     : !isContentImage(output);
   const normalizedContent = useMemo(() => normalizeContent(output), [output]);
 
-  const contents = mode !== "compact" ? input : input || functionCall;
+  const hasContent = normalizedContent.find((c) => {
+    if (c.type === "tool") {
+      for (const t of c.content) {
+        if (t.type === "text") {
+          if (t.text) {
+            return true;
+          }
+        } else {
+          return true;
+        }
+      }
+      return false;
+    } else {
+      return true;
+    }
+  });
 
+  const contents = mode !== "compact" ? input : input || functionCall;
   return (
     <div>
       {mode !== "compact" && (!view || view.title) ? (
@@ -99,9 +115,11 @@ export const ToolCallView: FC<ToolCallViewProps> = ({
             contents={contents}
             toolCallView={view}
           />
-          <ExpandablePanel collapse={collapse} border={true} lines={15}>
-            <MessageContent contents={normalizedContent} />
-          </ExpandablePanel>
+          {hasContent ? (
+            <ExpandablePanel collapse={collapse} border={true} lines={15}>
+              <MessageContent contents={normalizedContent} />
+            </ExpandablePanel>
+          ) : undefined}
         </div>
       </div>
     </div>

inspect_ai/_view/www/src/samples/chat/tools/ToolInput.tsx

@@ -64,7 +64,7 @@ export const ToolInput: FC<ToolInputProps> = memo((props) => {
         className={clsx(
           "source-code",
           "sourceCode",
-          `language-${highlightLanguage}`,
+          highlightLanguage ? `language-${highlightLanguage}` : undefined,
           styles.outputCode,
         )}
       >

inspect_ai/model/_openai.py

@@ -1,7 +1,9 @@
 import json
 import re
+from copy import copy
 from typing import Literal
 
+from openai import BadRequestError, OpenAIError
 from openai.types.chat import (
     ChatCompletion,
     ChatCompletionAssistantMessageParam,
@@ -26,7 +28,9 @@ from openai.types.chat.chat_completion import Choice, ChoiceLogprobs
 from openai.types.chat.chat_completion_message_tool_call import Function
 from openai.types.completion_usage import CompletionUsage
 from openai.types.shared_params.function_definition import FunctionDefinition
+from pydantic import JsonValue
 
+from inspect_ai._util.constants import BASE_64_DATA_REMOVED
 from inspect_ai._util.content import (
     Content,
     ContentAudio,
@@ -48,7 +52,16 @@ from ._chat_message import (
     ChatMessageTool,
     ChatMessageUser,
 )
-from ._model_output import ModelUsage, StopReason, as_stop_reason
+from ._model_output import ModelOutput, ModelUsage, StopReason, as_stop_reason
+
+
+class OpenAIResponseError(OpenAIError):
+    def __init__(self, code: str, message: str) -> None:
+        self.code = code
+        self.message = message
+
+    def __str__(self) -> str:
+        return f"{self.code}: {self.message}"
 
 
 def is_o_series(name: str) -> bool:
@@ -58,6 +71,10 @@ def is_o_series(name: str) -> bool:
         return not is_gpt(name) and bool(re.search(r"o\d+", name))
 
 
+def is_o1_pro(name: str) -> bool:
+    return "o1-pro" in name
+
+
 def is_o1_mini(name: str) -> bool:
     return "o1-mini" in name
 
@@ -320,6 +337,7 @@ def chat_messages_from_openai(
                 chat_messages.append(ChatMessageUser(content=content))
         elif message["role"] == "assistant":
             # resolve content
+            refusal: Literal[True] | None = None
             asst_content = message.get("content", None)
             if isinstance(asst_content, str):
                 result = parse_content_with_reasoning(asst_content)
@@ -336,6 +354,8 @@ def chat_messages_from_openai(
                     content = asst_content
             elif asst_content is None:
                 content = message.get("refusal", None) or ""
+                if content:
+                    refusal = True
             else:
                 content = []
                 for ac in asst_content:
@@ -348,7 +368,7 @@ def chat_messages_from_openai(
             )
             if reasoning is not None:
                 if isinstance(content, str):
-                    content = [ContentText(text=content)]
+                    content = [ContentText(text=content, refusal=refusal)]
                 else:
                     content.insert(0, ContentReasoning(reasoning=str(reasoning)))
 
@@ -437,7 +457,7 @@ def content_from_openai(
             )
         ]
     elif content["type"] == "refusal":
-        return [ContentText(text=content["refusal"])]
+        return [ContentText(text=content["refusal"], refusal=True)]
     else:
         content_type = content["type"]
         raise ValueError(f"Unexpected content type '{content_type}' in message.")
@@ -455,8 +475,10 @@ def chat_message_assistant_from_openai(
     if reasoning is not None:
         content: str | list[Content] = [
             ContentReasoning(reasoning=str(reasoning)),
-            ContentText(text=msg_content),
+            ContentText(text=msg_content, refusal=True if refusal else None),
         ]
+    elif refusal is not None:
+        content = [ContentText(text=msg_content, refusal=True)]
     else:
         content = msg_content
 
@@ -484,3 +506,44 @@ def chat_choices_from_openai(
         )
         for choice in choices
     ]
+
+
+def openai_handle_bad_request(
+    model_name: str, e: BadRequestError
+) -> ModelOutput | Exception:
+    # extract message
+    if isinstance(e.body, dict) and "message" in e.body.keys():
+        content = str(e.body.get("message"))
+    else:
+        content = e.message
+
+    # narrow stop_reason
+    stop_reason: StopReason | None = None
+    if e.code == "context_length_exceeded":
+        stop_reason = "model_length"
+    elif (
+        e.code == "invalid_prompt"  # seems to happen for o1/o3
+        or e.code == "content_policy_violation"  # seems to happen for vision
+        or e.code == "content_filter"  # seems to happen on azure
+    ):
+        stop_reason = "content_filter"
+
+    if stop_reason:
+        return ModelOutput.from_content(
+            model=model_name, content=content, stop_reason=stop_reason
+        )
+    else:
+        return e
+
+
+def openai_media_filter(key: JsonValue | None, value: JsonValue) -> JsonValue:
+    # remove images from raw api call
+    if key == "image_url" and isinstance(value, dict) and "url" in value:
+        url = str(value.get("url"))
+        if url.startswith("data:"):
+            value = copy(value)
+            value.update(url=BASE_64_DATA_REMOVED)
+    elif key == "input_audio" and isinstance(value, dict) and "data" in value:
+        value = copy(value)
+        value.update(data=BASE_64_DATA_REMOVED)
+    return value