@leostera/evalkit 0.5.0 → 0.6.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +70 -6
- package/README.md +10 -8
- package/dist/cli.mjs +1227 -258
- package/dist/dashboard/assets/index-CTJpd8q8.js +34 -0
- package/dist/dashboard/assets/index-YAabTN3X.css +2 -0
- package/dist/dashboard/index.html +2 -2
- package/dist/files-C-U0-PWW.mjs +0 -2
- package/dist/index-db-vbqEGFaU.mjs +0 -2
- package/dist/index.d.mts +1346 -2
- package/dist/index.d.mts.map +1 -1
- package/dist/index.mjs +2 -4
- package/dist/runner.d.mts +809 -11
- package/dist/runner.d.mts.map +1 -1
- package/dist/runner.mjs +2 -2
- package/dist/share-worker.d.mts +76 -1
- package/dist/share-worker.d.mts.map +1 -1
- package/dist/share-worker.mjs +0 -2
- package/dist/share.d.mts +149 -6
- package/dist/share.d.mts.map +1 -1
- package/dist/share.mjs +1 -1
- package/dist/skills/evalkit/SKILL.md +32 -0
- package/dist/skills/evalkit-agents/SKILL.md +16 -0
- package/dist/skills/evalkit-datasets/SKILL.md +17 -0
- package/dist/skills/evalkit-evals/SKILL.md +19 -0
- package/dist/skills/evalkit-judges/SKILL.md +17 -0
- package/dist/skills/evalkit-matrices/SKILL.md +14 -0
- package/dist/skills/evalkit-results/SKILL.md +14 -0
- package/dist/{src-DoPs9Yug.mjs → src-CKqOzeBi.mjs} +448 -116
- package/dist/{src-ButXPY8P.mjs → src-CZPufWN_.mjs} +419 -94
- package/dist/{src-DzZueFn3.mjs → src-CZnDXP1K.mjs} +46 -10
- package/package.json +3 -2
- package/dist/cli.mjs.map +0 -1
- package/dist/dashboard/assets/index-D9NtA46z.css +0 -2
- package/dist/dashboard/assets/index-DrBwLKJD.js +0 -30
- package/dist/files-C-U0-PWW.mjs.map +0 -1
- package/dist/index-db-HdZXelOS.d.mts +0 -142
- package/dist/index-db-HdZXelOS.d.mts.map +0 -1
- package/dist/index-db-vbqEGFaU.mjs.map +0 -1
- package/dist/index-lOw0fmak.d.mts +0 -1028
- package/dist/index-lOw0fmak.d.mts.map +0 -1
- package/dist/index.mjs.map +0 -1
- package/dist/share-worker.mjs.map +0 -1
- package/dist/src-ButXPY8P.mjs.map +0 -1
- package/dist/src-DoPs9Yug.mjs.map +0 -1
- package/dist/src-DzZueFn3.mjs.map +0 -1
package/dist/index.d.mts.map
CHANGED
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index.d.mts","names":[],"sources":["../../agents/src/index.ts"],"mappings":";;
|
|
1
|
+
{"version":3,"file":"index.d.mts","names":[],"sources":["../../core/src/identity.ts","../../core/src/dataset.ts","../../core/src/aut-events.ts","../../core/src/catalog.ts","../../core/src/canonical-json.ts","../../core/src/matrix.ts","../../core/src/config.ts","../../core/src/json.ts","../../core/src/definition-site.ts","../../core/src/schema.ts","../../core/src/trajectory.ts","../../core/src/index.ts","../../agents/src/index.ts"],"mappings":";;qBAEa,oBAAkB,OAAA;YACnB,sBAAsB,mBAAmB;;wBAGrC,SAAS;;qBAOZ,YAAU,OAAA,cAAA,OAAA;YAKX,cAAc,WAAW;YAEzB,YAAY,cAAc,eAAe,2BACxC;qBAEA,mBAAiB,OAAA,cAAA,OAAA;wBAMd,kBAAkB,cAAc,cAAc,MAAM,QAAK,OAAA,cAAA,OAAA;wBAWzD,YAAY,cAAc,cACxC,MAAM,OACN,MAAM,OACL,YAAY;wBAIC,iBAAiB,cAAc,cAC7C,eACA,eAAe;EACZ,MAAM;EAAO,MAAM;EAAM,KAAK,YAAY;;wBAY/B,aAAa,eAAe,eAAe,eAAe;;;;YCvD9D;;YAGA,6BAA6B;EAAe;;;;;;YAM5C,wBAAwB,SAAS,eAAe;;YAGhD;EACV,QAAQ;;EAER;;YAGU;;EAEV;EACA,QAAQ;;YAGE;EACV;EACA;EACA;EACA;;EAEA;;EAEA,YAAY;EACZ,QAAQ;;KAGL;EACH;EACA;;EAEA,aAAa;;EAEb;;YAGU,0BAA0B;EACpC,QAAQ;;EAER;;EAEA;EACA;EACA;;YAGU,0BAA0B,gBAAgB,kBACpD;;EAEE,SAAS,MAAM,uBAAuB,UAAU,QAAQ;;EAExD,SACE,QAAQ,SACR,OAAO,wBACJ,iBAAiB,QAAQ;EAC9B;EACA;EACA;;YAGQ;EAEN;EACA,QAAQ;EACR;EACA;;EAGA;EACA,SACE,MAAM,uBACH,iBAAiB,QAAQ;EAC9B,SACE,QAAQ,gBACR,OAAO,wBACJ,iBAAiB,QAAQ;;;YAIxB;WACD;WACA;WACA;WACA,YAAY;;WAEZ;WACA,QAAQ;;;wBAIH,oBAAoB,eAAe;;;;;;;wBAuBnC,cAAc,gBAAgB,gBAC5C,YAAY,0BAA0B,0BAA0B,WAC/D;wBAmFa,UAAU,iBAAiB,SAAS;;;;;;;YC/MxC;EACN;EAAiB;;EAEjB;EACA;EACA;EACA;;EAGA;EACA;EACA;EACA;EACA;;EAGA;EACA;EACA;EACA;EACA;;EAEA;EAAsB;EAAc;;EAEpC;EACA;EACA;EACA;EACA,QAAQ;;EAER;EAAmB;EAAkB;;EACrC;EAAe,OAAO;EAAe;;YAE/B;EACV;EACA;;EAEA;EACA;;;YAIU;EAGN;EACA,qBAAqB;EACrB,QAAQ;;;wBAIE,kBACd,OAAO,eACP,WAAW,OACV;;;;;wBA0Da,iBACd,QAAQ,YACR,cACA,WAAW,OACV;;;;;;;;;YCtGS;;EAEV;;YAGU;EACV;EACA;EACA;EACA;EACA,UAAU;IAAQ,MAAM;IAAkB;;;EAE1C;;;YAIU;EACN;EAAe;;EACf;EAAgB;;EAChB;;YAEM;EACV;EACA;EACA,YAAY;EACZ;EACA;;EAEA;IAAU;;IAAwB;;;EAElC;IAAU,OAAO;;IAAiB;;;EAElC;IAAU;IAAc;;EACxB;;YAGU;EACV;EACA,MAAM;EACN;EACA;;EAEA;EACA;;YAGU;EACN;EAAc;;EACd;EAAe,aAAa;;EAC5B;EAA0B;EAAc,YAAY;;;;EAEpD;EAAoB,QAAQ;;YAEtB;EACV;EACA;EACA;EACA;EACA,OAAO;;EAEP,aAAa;EACb,UAAU;EACV;EACA;IAAU;IAAgB;IAAoB;;EAC9C,YAAY;EACZ,SAAS;EACT,WAAW;;EAEX;;wBAGc,aACd,YAAY,gBACZ,OAAO,uBACP,UAAS,iBACR;;;;wBC3Fa,oBAAoB,OAAO;;;KCK/B;;EAEV;;EAEA,aAAa,SAAS,wBAAwB;;EAE9C,YAAY;;KAEF;EACV;;EAEA;EACA,MAAM;EACN,YAAY;;KAEF;EACV;EACA;EACA,gBAAgB;EAChB,YAAY,SAAS,wBAAwB;EAC7C,WAAW;;EAEX,iBAAiB;;EAEjB,mBAAmB;;KAET,aAAa;EACvB;EACA,MAAM,YAAY;EAClB,MAAM,YAAY,kBAAkB,iBAAiB;;wBAWvC,uBAAuB,UAAU,sBAC/C,YAAY,IACX,aAAa;;;KCjDJ;;EAEV;EACA;EACA;EACA;;IAEE;;;EAGF,iBAAiB;EACjB,WAAW;;EAEX,SAAS,KAAK;IAA0C;;EACxD;IAAc;IAAsB;IAAiB;;EACrD;EACA;EACA,QAAQ;;;KAIE,eAAe;KAEf;;EAEV;;;;;;EAMA,UAAU,sBAAsB,eAAe,QAAQ;;;wBAIzC,mBAAmB,UAAU,eAC3C,eAAe,IACd;;;;qBCrCU,qBAAqB;WAGrB;WACA;WAHO;EAClB,YACW,cACA;;;;;;;;wBAYG,OAAO,gBAAgB,gBAAiB;;wBA4CxC,aACd,gBACA;GACI,cAAc;;;;;;;;;;;KC3DR;EAAmB;EAAc;;wBAqB7B,eAAe,iBAAiB;;;;qBCvBnC,iBAAe,OAAA;qBAOf,qBAAmB,OAAA;;;;;qBAMnB,mBAAiB,OAAA;;;;;;qBAQjB,iBAAiB,OAAO,OAAO;qBAoB/B,iBAAe,OAAA;;;;;;;;;;;;;;qBAiBf,mBAAiB,OAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;qBAejB,qBAAmB,OAAA;;;;;;;;;;;;;;;;;;;;;;;;;;qBAyBnB,mBAAiB,OAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;qBAIjB,qBAAmB,OAAA;;;;;;;;;;;;;;;;;;;;;;;;;;qBAKnB,aAAW,OAAA;;;;;qBAMX,oBAAkB,OAAA;;;;;;;;;;;;;;qBAalB,mBAAiB,OAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;qBAYjB,wBAAsB,OAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;qBAgBtB,oBAAkB,OAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;qBASlB,qBAAmB,OAAA;;;;;qBAKnB,oBAAkB,OAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;qBASlB,kBAAgB,OAAA;;;;;;;;;;;;;;qBAmJhB,uBAAuB,OAAO,OAAO;qBAGrC,gBAAc,OAAA;;;;;;;wBCvUX,qBAAqB,mBAAmB;;;YCc5C;YACA,YACV,gBAAgB;GAAiB,cAAc;;YACrC;GAAgB,cAAc;;YAE9B;;EAEV;EACA;EACA;EACA;;YAGU;;;;;YAMA;EACV;EACA,gBAAgB;;YAGN,gBAAgB,QAAQ,OAAO,kBAAkB;YAEjD;EACV;EACA;EACA;;YAGU;EACV;EACA;EACA;;YAGU;EACN;EAAiB;;EAEjB;EACA;EACA,SAAS;EACT;;EAGA;EACA;EACA;EACA,WAAW;EACX;;EAGA;EACA;EACA;EACA,QAAQ;EACR;;EAEA;EAAsB;EAAc;;EAEpC;EACA;EACA;EACA;EACA,QAAQ;;EAER;EAAmB,SAAS;EAAW;;EACvC;EAAe,OAAO;EAAe;;YAE/B;EACN;EAAuB;;EAEvB;EACA;;EAEA;EACA;;EAEA;EAAmC;EAAc;;EACjD;EAAiC;EAAc;;EAE/C;EACA;EACA;EACA;;EAGA;EACA;EACA;EACA;;EAGA;EACA;EACA,OAAO;EACP;;EAEA;EAAwB;EAAgB;;EAExC;EACA;;EAEA;EACA;EACA;;EAGA;EACA;EACA,OAAO;EACP;;EAEA;EAAyB;;EACzB;EAAe,OAAO;EAAe;;YAE/B,mBACT;EAAa;MAAoB;EAAgB;;YAExC;;EAEV;;YAGU,qBAAqB;YACrB,qBAAqB;YAErB;EACV;EACA;EACA;EACA;EACA,UAAU;;EAEV,aAAa;EACb,UAAU;;EAEV,MAAM;;EAEN,WAAW;;EAEX,oBAAoB;;YAGV,WAAW;WACZ,WAAW;;WAEX,WAAW;EACpB,MAAM;IACJ,SAAS;IACT,UAAU;;IAEV,UAAU,OAAO,yBAAyB;MACxC,QAAQ,WAAW;;YAGb,WAAW;;;;;EAKrB,KAAK,SAAS,WAAW,eAAe;;EAExC,UAAU;;;qBAIC,wBAAwB;;wBAKrB,cAAc,OAAO,kBAAkB;;;;;wBAUvC,kBAAkB,UAAU,YAAY,OAAO,IAAI;YAKvD;EACV;EACA;;YAGU;EACV;EACA,aAAa;;YAGH;EACV;EACA,kBAAkB;IAAa;;EAC/B,6BAA6B,QAAQ;IAAY;;IAC/C;IACA;;EAEF;IACE;IACA;IACA;IACA,WAAW;IACX;IACA,SAAS;;EAEX;;YAGU;EACV;EACA;EACA;IAAY;IAAc,YAAY;;;YAE5B,iBACV,WAAW,YAAY,cAAc;;wBAGvB,eACd,UAAU,iCACT;wBAMa,KAAK,kBAAkB;wBAIvB,MAAM,aAAa,aAAa;YAIpC;EACV,QAAQ;EACR,QAAQ;;EAER,SAAS;;;YAIC;EAAiB;EAAe;;YAEhC,sBAER;EAEE;EACA;EACA;;EAEA;;EAEA,QAAQ;;;;;;wBAOE,KAAK,iBAAiB;YAK1B;EACV;EACA;EACA;EACA;EACA;EACA,WAAW;EACX,QAAQ;EACR;EACA,QAAQ;;EAER;;YAGU;EACV,WAAW;EACX,WAAW;;YAGD;EACV,SAAS;EACT;IAAc,iBAAiB;;EAC/B,WAAW;;EAEX,OAAO;;EAEP,SAAS;;EAET,MAAM;;YAGI;EACV;EACA;EACA;;EAEA;EACA,IACE,SAAS,2BACE,aAAa,kBAAkB;;YAGlC;EACV;EACA;EACA;;EAEA;;YAGU,cAAc,kBAAkB;wBAE5B,UACd,cACA,KAAK,wBACL,UAAS,KAAK,iDACb;wBAgBa,MACd,cACA,SAAS,KAAK,2CACb;YAMS;EACV;;EAEA;;EAEA;;YAGU,eAAe,eAAe,aAAa;EACrD;EACA;EACA,OAAO;;EAEP,QAAQ;;EAER,oBAAoB;EACpB,YAAY;EACZ,SAAS;EACT,SAAS;;EAET,WAAW;;wBAGG,iBAAiB,UAAU,gBAAgB,YAAY,IAAI;YAyC/D,UACV,wBAAwB,4BAA4B;EAEpD;EACA;EACA,OAAO;;wBAGO,kBAAkB,eAAe,WAC/C,OAAO,SACN;YAKS,mBAAmB,iBAAiB,YAAY;wBAE5C,YAAY;EAAS;;YAkBzB;;EAEV,gBAAgB;EAChB,iBAAiB;EACjB,mBAAmB;EACnB,IAAI,cAAc;EAClB,SAAS,cAAc;EACvB,YAAY;EACZ,iBAAiB;;EAEjB,QAAQ,UAAU,iBAAiB;;YAGzB;EACV;EACA;EACA;;YAGU;EACV;EACA;EACA;;;;;;wBAOc,oBACR,gCAAgC,oBACtC,eAAe,iBAAiB;YAuEtB;;;;;YAMA;;EAEV;;EAEA;;EAEA;EACA;EACA;;EAEA;EACA;IAAO;IAAkB;IAAmB;IAAc;;;;YAIhD;EACV,SAAS;EACT,OAAO;EACP,QAAQ;;;YAIE;EAAe;EAAa;;YAE5B;EACV;EACA;EACA,QAAQ;EACR;EACA;EACA,aAAa;EACb;IAAW;IAAY;;EACvB,MAAM;EACN;;EAEA,SAAS;;EAET,UAAU;;YAGA;EACV;EACA;EACA,QAAQ;EACR;EACA,UAAU;EACV;EACA;EACA,aAAa;EACb;IAAW;IAAY;;EACvB,MAAM;EACN;;EAEA,WAAW;;YAGD;EACV;EACA,YAAY;EACZ;;EAEA;;YAGU;EACV;EACA;EACA;EACA;EACA;EACA;EACA;EACA,WAAW;EACX,QAAQ;EACR;EACA,QAAQ;EACR;IAAoB;IAAoB;;;YAG9B;EACV,SAAS;EACT,cAAc;EACd;EACA;EACA;;YAGU;;EAEV;EACA;EACA;;YAGU;EACV,QAAQ;EACR;EACA;EACA,UAAU;EACV,YAAY;EACZ,QAAQ;;YAGE;EACV,QAAQ;EACR;EACA;EACA;EACA;EACA;;EAEA;EACA,QAAQ;;YAGE;EACV,YAAY,OAAO,kBAAkB;EACrC,YAAY,SAAS,eAAe;;EAEpC,cAAc,cAAc,MAAM,aAAa;EAC/C,SAAS,SAAS,eAAe;;YAGvB;WACD;EACT,WAAW,UAAU,gBAAgB,QAAQ;EAC7C,SAAS,SAAS,aAAa;;YAGrB;EACV,SACE,UAAU,aACV;IAAY,UAAU;MACrB,QAAQ;;YAGD;EACV;EACA,QAAQ;EACR;EACA,UAAU;EACV;EACA;EACA,QAAQ;EACR;EACA;EACA,UAAU;EACV,QAAQ;;YAGE;EACV;EACA;EACA;EACA;EACA;;YAGU,YAAY;EACtB;EACA;EACA;EACA;EACA,mBAAmB;;EAEnB,SAAS;;wBAYK,YAAY,iBAAiB;;;KCxuBjC;EACV;;EAEA;;;wBAIc,QAAQ,UAAS,eAAoB"}
|
package/dist/index.mjs
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { A as
|
|
1
|
+
import { $ as resourceUriSchema, A as TrialManifestSchema, B as NotJsonError, C as RunManifestSchema, D as RunSummarySchema, E as RunStatusSchema, F as normalizeAutEvent, G as relativeDatasetPath, H as toJsonObject, I as turnResultEvents, J as ResourceUriSchema, K as definitionSite, L as defineEvalMatrix, M as TrialScoringSchema, N as TrialSummarySchema, O as ScoreResultSchema, P as UsageSchema, Q as resourceUri, R as canonicalParameters, S as RecordedErrorSchema, T as RunSourceSchema, U as defineDataset, V as toJson, W as isDataset, X as authorId, Y as UuidSchema, Z as parseResourceUri, _ as ArtifactEntrySchema, a as defineAgent, b as JsonValueSchema, c as expectToolCall, d as recordError, et as resourceUuid, f as registerEvals, g as ApiErrorSchema, h as parseTrajectoryJsonl, i as authoringId, j as TrialMetadataSchema, k as TrajectoryEventSchema, l as judge, m as user, n as agent, o as defineEval, p as skip, q as ResourceKindSchema, r as agentRuntimes, s as defineSuite, t as DEFAULT_AGENT_RUNTIMES, u as predicate, v as AutIdentitySchema, w as RunMetadataSchema, x as JudgeRunInfoSchema, y as CheckpointResultSchema, z as catalogEntry } from "./src-CZPufWN_.mjs";
|
|
2
2
|
//#region ../core/src/config.ts
|
|
3
3
|
/** Pure, typed project configuration. Loading/discovery belongs to the CLI. */
|
|
4
4
|
function defineConfig(configuration) {
|
|
@@ -74,6 +74,4 @@ function piAgent(options = {}) {
|
|
|
74
74
|
});
|
|
75
75
|
}
|
|
76
76
|
//#endregion
|
|
77
|
-
export { ApiErrorSchema, ArtifactEntrySchema, AutIdentitySchema, CheckpointResultSchema, JsonValueSchema, JudgeRunInfoSchema, RecordedErrorSchema, ResourceKindSchema, ResourceUriSchema, RunManifestSchema, RunMetadataSchema, RunSourceSchema, RunStatusSchema, RunSummarySchema, ScoreResultSchema, TrajectoryEventSchema, TrialManifestSchema, TrialMetadataSchema, TrialScoringSchema, TrialSummarySchema, UsageSchema, UuidSchema, agent, authorId, authoringId, canonicalParameters, defineAgent, defineConfig, defineEval, defineEvalMatrix, defineSuite,
|
|
78
|
-
|
|
79
|
-
//# sourceMappingURL=index.mjs.map
|
|
77
|
+
export { ApiErrorSchema, ArtifactEntrySchema, AutIdentitySchema, CheckpointResultSchema, DEFAULT_AGENT_RUNTIMES, JsonValueSchema, JudgeRunInfoSchema, NotJsonError, RecordedErrorSchema, ResourceKindSchema, ResourceUriSchema, RunManifestSchema, RunMetadataSchema, RunSourceSchema, RunStatusSchema, RunSummarySchema, ScoreResultSchema, TrajectoryEventSchema, TrialManifestSchema, TrialMetadataSchema, TrialScoringSchema, TrialSummarySchema, UsageSchema, UuidSchema, agent, agentRuntimes, authorId, authoringId, canonicalParameters, catalogEntry, defineAgent, defineConfig, defineDataset, defineEval, defineEvalMatrix, defineSuite, definitionSite, expectToolCall, isDataset, judge, normalizeAutEvent, parseResourceUri, parseTrajectoryJsonl, piAgent, predicate, recordError, registerEvals, relativeDatasetPath, resourceUri, resourceUriSchema, resourceUuid, skip, toJson, toJsonObject, turnResultEvents, user };
|