inspect-ai 0.3.82__py3-none-any.whl → 0.3.83__py3-none-any.whl

inspect_ai/_view/www/src/components/Modal.tsx

@@ -0,0 +1,77 @@
+import clsx from "clsx";
+import { FC, ReactNode } from "react";
+import styles from "./Modal.module.css";
+
+interface ModalProps {
+  id: string;
+  showing: boolean;
+  setShowing: (showing: boolean) => void;
+  title?: string;
+  children: ReactNode;
+  className?: string | string[];
+}
+
+export const Modal: FC<ModalProps> = ({
+  id,
+  title,
+  showing,
+  setShowing,
+  children,
+  className,
+}) => {
+  return (
+    <>
+      {showing && (
+        <div className={styles.backdrop} onClick={() => setShowing(false)} />
+      )}
+      <div
+        id={id}
+        className={clsx("modal", "fade", showing ? "show" : "", className)}
+        tabIndex={-1}
+        style={{ display: showing ? "block" : "none" }}
+      >
+        <div className={clsx("modal-dialog", styles.modal)}>
+          <div className="modal-content">
+            <div className={clsx("modal-header", styles.header)}>
+              <div
+                className={clsx(
+                  "modal-title",
+                  "text-size-base",
+                  styles.modalTitle,
+                )}
+              >
+                {title}
+              </div>
+              <button
+                type="button"
+                className={clsx(
+                  "btn-close",
+                  "text-size-smaller",
+                  styles.btnClose,
+                )}
+                data-bs-dismiss="modal"
+                aria-label="Close"
+                onClick={() => {
+                  setShowing(!showing);
+                }}
+              ></button>
+            </div>
+            <div className="modal-body">{children}</div>
+            <div className="modal-footer">
+              <button
+                type="button"
+                className="btn btn-secondary"
+                data-bs-dismiss="modal"
+                onClick={() => {
+                  setShowing(!showing);
+                }}
+              >
+                Close
+              </button>
+            </div>
+          </div>
+        </div>
+      </div>
+    </>
+  );
+};

inspect_ai/_view/www/src/plan/DetailStep.module.css

@@ -7,3 +7,7 @@
   margin-top: 0.2rem;
   margin-bottom: 0.3rem;
 }
+
+.metadata {
+  margin-bottom: 0.75em;
+}

inspect_ai/_view/www/src/plan/DetailStep.tsx

@@ -1,7 +1,7 @@
 import clsx from "clsx";
 import { FC } from "react";
-import { MetaDataView } from "../metadata/MetaDataView";
-import styles from "./DatasetDetailView.module.css";
+import { MetaDataGrid } from "../metadata/MetaDataGrid";
+import styles from "./DetailStep.module.css";
 
 interface DetailStepProps {
   icon?: string;
@@ -22,7 +22,10 @@ export const DetailStep: FC<DetailStepProps> = ({
       {iconHtml} {name}
       <div className={styles.container}>
         {params ? (
-          <MetaDataView entries={params} className={"text-size-small"} />
+          <MetaDataGrid
+            entries={params}
+            className={clsx("text-size-small", styles.metadata)}
+          />
         ) : (
           ""
         )}

inspect_ai/_view/www/src/plan/SolverDetailView.module.css

@@ -1,6 +1,7 @@
 .container {
   display: flex;
-  flex-direction: columns;
+  flex-direction: row;
+  flex-wrap: wrap;
 }
 
 .item {

inspect_ai/_view/www/src/samples/InlineSampleDisplay.tsx

@@ -25,8 +25,15 @@ export const InlineSampleDisplay: FC<InlineSampleDisplayProps> = ({
   // Sample hooks
   const sampleData = useSampleData();
   const loadSample = useStore((state) => state.sampleActions.loadSample);
+  const pollSample = useStore((state) => state.sampleActions.pollSample);
   const logSelection = useLogSelection();
 
+  useEffect(() => {
+    if (sampleData.running && logSelection.logFile && logSelection.sample) {
+      pollSample(logSelection.logFile, logSelection.sample);
+    }
+  }, []);
+
   // Sample Loading
   const prevCompleted = usePrevious(
     logSelection.sample?.completed !== undefined

inspect_ai/_view/www/src/samples/SampleDialog.tsx

@@ -39,8 +39,15 @@ export const SampleDialog: FC<SampleDialogProps> = ({
   // Sample hooks
   const sampleData = useSampleData();
   const loadSample = useStore((state) => state.sampleActions.loadSample);
+  const pollSample = useStore((state) => state.sampleActions.pollSample);
   const logSelection = useLogSelection();
 
+  useEffect(() => {
+    if (sampleData.running && logSelection.logFile && logSelection.sample) {
+      pollSample(logSelection.logFile, logSelection.sample);
+    }
+  }, []);
+
   // Load sample
   const prevCompleted = usePrevious(
     logSelection.sample?.completed !== undefined

inspect_ai/_view/www/src/samples/SampleDisplay.tsx

@@ -7,7 +7,6 @@ import { isVscode } from "../utils/vscode";
 import { ApplicationIcons } from "../appearance/icons";
 import { ANSIDisplay } from "../components/AnsiDisplay";
 import { ToolButton } from "../components/ToolButton";
-import { SampleScoreView } from "./scores/SampleScoreView";
 
 import clsx from "clsx";
 import {
@@ -40,6 +39,7 @@ import { ChatViewVirtualList } from "./chat/ChatViewVirtualList";
 import { messagesFromEvents } from "./chat/messages";
 import styles from "./SampleDisplay.module.css";
 import { SampleSummaryView } from "./SampleSummaryView";
+import { SampleScoresView } from "./scores/SampleScoresView";
 import { TranscriptVirtualList } from "./transcript/TranscriptView";
 
 interface SampleDisplayProps {
@@ -92,7 +92,6 @@ export const SampleDisplay: FC<SampleDisplayProps> = ({
     return false;
   };
 
-  const scorerNames = Object.keys(sample?.scores || {});
   const sampleMetadatas = metadataViewsForSample(`${baseId}-${id}`, sample);
 
   const tabsetId = `task-sample-details-tab-${id}`;
@@ -166,38 +165,16 @@ export const SampleDisplay: FC<SampleDisplayProps> = ({
             running={running}
           />
         </TabPanel>
-        {sample && scorerNames.length === 1 ? (
-          <TabPanel
-            key={kSampleScoringTabId}
-            id={kSampleScoringTabId}
-            className="sample-tab"
-            title="Scoring"
-            onSelected={onSelectedTab}
-            selected={selectedTab === kSampleScoringTabId}
-          >
-            <SampleScoreView sample={sample} scorer={scorerNames[0]} />
-          </TabPanel>
-        ) : (
-          <>
-            {sample
-              ? Object.keys(sample?.scores || {}).map((scorer) => {
-                  const tabId = `score-${scorer}`;
-                  return (
-                    <TabPanel
-                      key={tabId}
-                      id={tabId}
-                      className="sample-tab"
-                      title={scorer}
-                      onSelected={onSelectedTab}
-                      selected={selectedTab === tabId}
-                    >
-                      <SampleScoreView sample={sample} scorer={scorer} />
-                    </TabPanel>
-                  );
-                })
-              : undefined}
-          </>
-        )}
+        <TabPanel
+          key={kSampleScoringTabId}
+          id={kSampleScoringTabId}
+          className="sample-tab"
+          title="Scoring"
+          onSelected={onSelectedTab}
+          selected={selectedTab === kSampleScoringTabId}
+        >
+          <SampleScoresView sample={sample} />
+        </TabPanel>
         <TabPanel
           id={kSampleMetdataTabId}
           className={clsx("sample-tab")}

inspect_ai/_view/www/src/samples/SampleSummaryView.module.css

@@ -19,6 +19,11 @@
   justify-content: center;
 }
 
+.centerValue {
+  display: flex;
+  align-items: center;
+}
+
 .wrap {
   word-wrap: anywhere;
 }
@@ -29,4 +34,5 @@
 
 .value {
   flex-direction: column;
+  padding-top: 0.1em;
 }

inspect_ai/_view/www/src/samples/SampleSummaryView.tsx

@@ -123,7 +123,7 @@ export const SampleSummaryView: FC<SampleSummaryViewProps> = ({
 
   columns.push({
     label: "Input",
-    value: fields.input,
+    value: <MarkdownDiv markdown={fields.input.join(" ")} />,
     size: `${input}fr`,
     clamp: true,
   });
@@ -233,7 +233,7 @@ export const SampleSummaryView: FC<SampleSummaryViewProps> = ({
               styles.value,
               styles.wrap,
               col.clamp ? "three-line-clamp" : undefined,
-              col.center ? styles.centerLabel : undefined,
+              col.center ? styles.centerValue : undefined,
             )}
           >
             {col.value}

inspect_ai/_view/www/src/samples/SamplesTools.tsx

@@ -44,3 +44,15 @@ export const SampleTools: FC<SampleToolsProps> = ({ samples }) => {
     </Fragment>
   );
 };
+
+interface ScoreFilterToolsProps {}
+
+export const ScoreFilterTools: FC<ScoreFilterToolsProps> = () => {
+  const scores = useScores();
+  const score = useScore();
+  const setScore = useStore((state) => state.logActions.setScore);
+  if (scores.length <= 1) {
+    return undefined;
+  }
+  return <SelectScorer scores={scores} score={score} setScore={setScore} />;
+};

inspect_ai/_view/www/src/samples/chat/MessageContent.tsx

@@ -52,6 +52,7 @@ export const MessageContent: FC<MessageContentProps> = ({ contents }) => {
           {
             type: "text",
             text: content,
+            refusal: null,
           },
           index === contents.length - 1,
         );
@@ -75,6 +76,7 @@ export const MessageContent: FC<MessageContentProps> = ({ contents }) => {
     const contentText: ContentText = {
       type: "text",
       text: contents,
+      refusal: null,
     };
     return messageRenderers["text"].render(
       "text-message-content",

inspect_ai/_view/www/src/samples/chat/MessageContents.tsx

@@ -101,6 +101,7 @@ const resolveToolMessage = (toolMessage?: ChatMessageTool): ContentTool[] => {
           {
             type: "text",
             text: content,
+            refusal: null,
           },
         ],
       },
@@ -115,6 +116,7 @@ const resolveToolMessage = (toolMessage?: ChatMessageTool): ContentTool[] => {
               {
                 type: "text",
                 text: con,
+                refusal: null,
               },
             ],
           } as ContentTool;

inspect_ai/_view/www/src/samples/chat/messages.ts

@@ -70,6 +70,7 @@ export const resolveMessages = (messages: Messages) => {
     role: "system",
     content: systemContent,
     source: "input",
+    internal: null,
   };
 
   // Converge them
@@ -120,6 +121,7 @@ const normalizeContent = (
     return {
       type: "text",
       text: content,
+      refusal: null,
     };
   } else {
     return content;
@@ -151,7 +153,7 @@ export const messagesFromEvents = (runningEvents: Events): Messages => {
       }
     });
 
-  if (messages.entries.length > 0) {
+  if (messages.size > 0) {
     return messages.values().toArray();
   } else {
     return [];

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

@@ -175,6 +175,7 @@ const normalizeContent = (
           {
             type: "text",
             text: String(output),
+            refusal: null,
           },
         ],
       },

inspect_ai/_view/www/src/samples/descriptor/samplesDescriptor.tsx

@@ -47,7 +47,13 @@ export const createEvalDescriptor = (
       sample.scores[scoreLabel.scorer] &&
       sample.scores[scoreLabel.scorer].value
     ) {
-      return sample.scores[scoreLabel.scorer].value;
+      if (typeof sample.scores[scoreLabel.scorer].value === "object") {
+        return (
+          sample.scores[scoreLabel.scorer].value as Record<string, Value2>
+        )[scoreLabel.name];
+      } else {
+        return sample.scores[scoreLabel.scorer].value;
+      }
     } else if (sample.scores[scoreLabel.name]) {
       return sample.scores[scoreLabel.name].value;
     } else {
@@ -162,7 +168,7 @@ export const createEvalDescriptor = (
       return "null";
     } else if (score === undefined) {
       return "";
-    } else if (score && descriptor && descriptor.render) {
+    } else if (descriptor && descriptor.render) {
       return descriptor.render(score);
     } else {
       return <span>{String(score)}</span>;
@@ -328,7 +334,7 @@ export const createSamplesDescriptor = (
     answer: Math.min(sizes[2], 300),
     limit: Math.min(sizes[3], 50),
     id: Math.min(sizes[4], 10),
-    score: Math.min(sizes[4], 30),
+    score: Math.min(sizes[5], 30),
   };
   const base =
     maxSizes.input +

inspect_ai/_view/www/src/samples/descriptor/score/BooleanScoreDescriptor.module.css

@@ -1,11 +1,11 @@
 .circle {
   font-family: "Consola Regular";
-  width: 20px;
-  height: 20px;
+  width: 40px;
+  height: 30px;
   display: inline-flex;
   justify-content: center;
   align-items: center;
-  border-radius: 50%;
+  border-radius: 15px;
   padding-top: 1px;
 }
 

inspect_ai/_view/www/src/samples/descriptor/score/BooleanScoreDescriptor.tsx

@@ -15,7 +15,7 @@ export const booleanScoreDescriptor = (): ScoreDescriptor => {
           className={clsx(
             styles.circle,
             "text-size-small",
-            score ? "green" : "red",
+            score ? styles.green : styles.red,
           )}
         >
           {String(score)}

inspect_ai/_view/www/src/samples/descriptor/score/ObjectScoreDescriptor.module.css

@@ -1,8 +1,8 @@
 .container {
-  display: flex;
-  flex-direction: column;
-  align-items: center;
-  margin-left: 0.5rem;
+  display: grid;
+  grid-template-columns: auto auto;
+  grid-auto-rows: auto;
+  column-gap: 1rem;
 }
 
 .padded {

inspect_ai/_view/www/src/samples/descriptor/score/ObjectScoreDescriptor.tsx

@@ -35,7 +35,7 @@ export const objectScoreDescriptor = (values: Value2[]): ScoreDescriptor => {
 
       const scores: JSX.Element[] = [];
       const keys = Object.keys(score);
-      keys.forEach((key, index) => {
+      keys.forEach((key) => {
         if (typeof score !== "object" || Array.isArray(score)) {
           throw new Error(
             "Unexpected us of object score descriptor for non-score object",
@@ -50,23 +50,22 @@ export const objectScoreDescriptor = (values: Value2[]): ScoreDescriptor => {
                   : parseFloat(value === true ? "1" : value),
               )
             : String(value);
+
         scores.push(
-          <div
-            key={`score-value-${index}`}
-            className={clsx(
-              styles.container,
-              index + 1 < keys.length ? styles.padded : undefined,
-            )}
-          >
+          <>
             <div className={clsx(styles.key, "text-size-smaller")}>{key}</div>
-            <div className={clsx(styles.value, "text-size-large")}>
+            <div className={clsx(styles.value, "text-size-base")}>
               {formattedValue}
             </div>
-          </div>,
+          </>,
         );
       });
 
-      return scores;
+      return (
+        <div key={`score-value`} className={clsx(styles.container)}>
+          {scores}
+        </div>
+      );
     },
   };
 };

inspect_ai/_view/www/src/samples/list/SampleFooter.module.css

@@ -23,5 +23,6 @@
 }
 
 .label {
-  margin-top: -4px;
+  margin-left: 0.1em;
+  margin-top: -3px;
 }

inspect_ai/_view/www/src/samples/list/SampleFooter.tsx

@@ -1,5 +1,6 @@
 interface SampleFooterProps {
   sampleCount: number;
+  totalSampleCount: number;
   running: boolean;
 }
 
@@ -9,6 +10,7 @@ import styles from "./SampleFooter.module.css";
 
 export const SampleFooter: FC<SampleFooterProps> = ({
   sampleCount,
+  totalSampleCount,
   running,
 }) => {
   return (
@@ -28,7 +30,11 @@ export const SampleFooter: FC<SampleFooterProps> = ({
           </div>
         ) : undefined}
       </div>
-      <div>{sampleCount} Samples</div>
+      <div>
+        {sampleCount < totalSampleCount
+          ? `${sampleCount} / ${totalSampleCount} Samples`
+          : `${sampleCount} Samples`}
+      </div>
     </div>
   );
 };

inspect_ai/_view/www/src/samples/list/SampleList.tsx

@@ -29,6 +29,7 @@ const kSeparatorHeight = 24;
 
 interface SampleListProps {
   items: ListItem[];
+  totalItemCount: number;
   running: boolean;
   nextSample: () => void;
   prevSample: () => void;
@@ -37,9 +38,12 @@ interface SampleListProps {
   listHandle: RefObject<VirtuosoHandle | null>;
 }
 
+export const kSampleFollowProp = "sample-list";
+
 export const SampleList: FC<SampleListProps> = memo((props) => {
   const {
     items,
+    totalItemCount,
     running,
     nextSample,
     prevSample,
@@ -57,9 +61,13 @@ export const SampleList: FC<SampleListProps> = memo((props) => {
     (state) => state.log.selectedSampleIndex,
   );
   const samplesDescriptor = useSampleDescriptor();
-  const [followOutput, setFollowOutput] = useProperty("sample-list", "follow", {
-    defaultValue: false,
-  });
+  const [followOutput, setFollowOutput] = useProperty(
+    kSampleFollowProp,
+    "follow",
+    {
+      defaultValue: !!running,
+    },
+  );
 
   // Track whether we were previously running so we can
   // decide whether to pop up to the top
@@ -84,13 +92,15 @@ export const SampleList: FC<SampleListProps> = memo((props) => {
     prevRunningRef.current = running;
   }, [running, followOutput, listHandle]);
 
+  const loaded = useRef(false);
   const handleAtBottomStateChange = useCallback(
     (atBottom: boolean) => {
-      if (running) {
+      if (loaded.current && running) {
         setFollowOutput(atBottom);
       }
+      loaded.current = true;
     },
-    [running, setFollowOutput],
+    [running, setFollowOutput, followOutput],
   );
 
   const onkeydown = useCallback(
@@ -148,7 +158,7 @@ export const SampleList: FC<SampleListProps> = memo((props) => {
         return null;
       }
     },
-    [showSample],
+    [showSample, gridColumnsTemplate],
   );
 
   const { input, limit, answer, target } = gridColumns(samplesDescriptor);
@@ -210,8 +220,11 @@ export const SampleList: FC<SampleListProps> = memo((props) => {
         data={items}
         defaultItemHeight={50}
         itemContent={renderRow}
-        followOutput={followOutput}
+        followOutput={(_atBottom: boolean) => {
+          return followOutput;
+        }}
         atBottomStateChange={handleAtBottomStateChange}
+        atBottomThreshold={30}
         increaseViewportBy={{ top: 300, bottom: 300 }}
         overscan={{
           main: 10,
@@ -223,7 +236,11 @@ export const SampleList: FC<SampleListProps> = memo((props) => {
         isScrolling={isScrolling}
         restoreStateFrom={getRestoreState()}
       />
-      <SampleFooter sampleCount={sampleCount} running={running} />
+      <SampleFooter
+        sampleCount={sampleCount}
+        totalSampleCount={totalItemCount}
+        running={running}
+      />
     </div>
   );
 });

inspect_ai/_view/www/src/samples/list/SampleRow.tsx

@@ -69,7 +69,7 @@ export const SampleRow: FC<SampleRowProps> = ({
           styles.wrapAnywhere,
         )}
       >
-        {inputString(sample.input).join(" ")}
+        <MarkdownDiv markdown={inputString(sample.input).join(" ")} />
       </div>
       <div className={clsx("sample-target", "three-line-clamp", styles.cell)}>
         <MarkdownDiv

inspect_ai/_view/www/src/samples/scores/SampleScores.tsx

@@ -1,8 +1,7 @@
-import { FC, Fragment } from "react";
+import { FC } from "react";
 import { SampleSummary } from "../../api/types";
 
-import { useSampleDescriptor } from "../../state/hooks";
-import styles from "./SampleScores.module.css";
+import { getScoreDescriptorForValues } from "../descriptor/score/ScoreDescriptor";
 
 interface SampleScoresProps {
   sample: SampleSummary;
@@ -10,24 +9,14 @@ interface SampleScoresProps {
 }
 
 export const SampleScores: FC<SampleScoresProps> = ({ sample, scorer }) => {
-  const samplesDescriptor = useSampleDescriptor();
-  const scores = scorer
-    ? samplesDescriptor?.evalDescriptor
-        .scorerDescriptor(sample, { scorer, name: scorer })
-        .scores()
-    : samplesDescriptor?.selectedScorerDescriptor(sample)?.scores();
-
-  if (scores?.length === 1) {
-    return scores[0].rendered();
-  } else {
-    const rows = scores?.map((score) => {
-      return (
-        <Fragment>
-          <div style={{ opacity: "0.7" }}>{score.name}</div>
-          <div>{score.rendered()}</div>
-        </Fragment>
-      );
-    });
-    return <div className={styles.grid}>{rows}</div>;
+  const scoreData = sample.scores?.[scorer];
+  if (!scoreData) {
+    return undefined;
   }
+
+  const scorerDescriptor = getScoreDescriptorForValues(
+    [scoreData.value],
+    [typeof scoreData.value],
+  );
+  return scorerDescriptor?.render(scoreData.value);
 };