@mcoda/mswarm 0.1.90 → 0.1.94
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/dist/codali-executor.d.ts +107 -6
- package/dist/codali-executor.d.ts.map +1 -1
- package/dist/codali-executor.js +500 -26
- package/dist/codali-executor.js.map +1 -1
- package/dist/runtime.d.ts +4 -2
- package/dist/runtime.d.ts.map +1 -1
- package/dist/runtime.js +18 -5
- package/dist/runtime.js.map +1 -1
- package/dist/vendor/codali/cli/DatasetCommand.d.ts +38 -0
- package/dist/vendor/codali/cli/DatasetCommand.d.ts.map +1 -0
- package/dist/vendor/codali/cli/DatasetCommand.js +959 -0
- package/dist/vendor/codali/cli/EvalCommand.d.ts +3 -0
- package/dist/vendor/codali/cli/EvalCommand.d.ts.map +1 -1
- package/dist/vendor/codali/cli/EvalCommand.js +55 -1
- package/dist/vendor/codali/cli/ImprovementCommand.d.ts +97 -0
- package/dist/vendor/codali/cli/ImprovementCommand.d.ts.map +1 -0
- package/dist/vendor/codali/cli/ImprovementCommand.js +1879 -0
- package/dist/vendor/codali/cli.d.ts.map +1 -1
- package/dist/vendor/codali/cli.js +15 -0
- package/dist/vendor/codali/dataset-cli.d.ts +3 -0
- package/dist/vendor/codali/dataset-cli.d.ts.map +1 -0
- package/dist/vendor/codali/dataset-cli.js +30 -0
- package/dist/vendor/codali/eval/CodaliGatewayLiveHarness.d.ts +90 -1
- package/dist/vendor/codali/eval/CodaliGatewayLiveHarness.d.ts.map +1 -1
- package/dist/vendor/codali/eval/CodaliGatewayLiveHarness.js +443 -9
- package/dist/vendor/codali/eval/GatewayDatasetEval.d.ts +70 -0
- package/dist/vendor/codali/eval/GatewayDatasetEval.d.ts.map +1 -0
- package/dist/vendor/codali/eval/GatewayDatasetEval.js +477 -0
- package/dist/vendor/codali/eval/GatewayEvalSuite.d.ts +52 -0
- package/dist/vendor/codali/eval/GatewayEvalSuite.d.ts.map +1 -1
- package/dist/vendor/codali/eval/GatewayEvalSuite.js +135 -1
- package/dist/vendor/codali/gateway/CodaliGateway.d.ts +4 -0
- package/dist/vendor/codali/gateway/CodaliGateway.d.ts.map +1 -1
- package/dist/vendor/codali/gateway/CodaliGateway.js +37 -1
- package/dist/vendor/codali/gateway/CodaliGatewaySchemas.d.ts.map +1 -1
- package/dist/vendor/codali/gateway/CodaliGatewaySchemas.js +0 -4
- package/dist/vendor/codali/gateway/CodaliGatewayStore.d.ts.map +1 -1
- package/dist/vendor/codali/gateway/CodaliGatewayStore.js +1 -0
- package/dist/vendor/codali/gateway/GatewayPlanner.d.ts.map +1 -1
- package/dist/vendor/codali/gateway/GatewayPlanner.js +355 -41
- package/dist/vendor/codali/gateway/GatewayStateMachine.d.ts +1 -0
- package/dist/vendor/codali/gateway/GatewayStateMachine.d.ts.map +1 -1
- package/dist/vendor/codali/gateway/GatewayStateMachine.js +221 -5
- package/dist/vendor/codali/improvement/CandidateReleaseBuilder.d.ts +201 -0
- package/dist/vendor/codali/improvement/CandidateReleaseBuilder.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/CandidateReleaseBuilder.js +1175 -0
- package/dist/vendor/codali/improvement/DatasetEligibilityGate.d.ts +75 -0
- package/dist/vendor/codali/improvement/DatasetEligibilityGate.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/DatasetEligibilityGate.js +497 -0
- package/dist/vendor/codali/improvement/DatasetExportManifestReader.d.ts +88 -0
- package/dist/vendor/codali/improvement/DatasetExportManifestReader.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/DatasetExportManifestReader.js +333 -0
- package/dist/vendor/codali/improvement/DocdexRetrievalCandidateBuilder.d.ts +192 -0
- package/dist/vendor/codali/improvement/DocdexRetrievalCandidateBuilder.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/DocdexRetrievalCandidateBuilder.js +625 -0
- package/dist/vendor/codali/improvement/EvalReplayCandidateBuilder.d.ts +173 -0
- package/dist/vendor/codali/improvement/EvalReplayCandidateBuilder.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/EvalReplayCandidateBuilder.js +363 -0
- package/dist/vendor/codali/improvement/FineTuneJobPlanner.d.ts +281 -0
- package/dist/vendor/codali/improvement/FineTuneJobPlanner.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/FineTuneJobPlanner.js +823 -0
- package/dist/vendor/codali/improvement/ImprovementEvalRunner.d.ts +50 -0
- package/dist/vendor/codali/improvement/ImprovementEvalRunner.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/ImprovementEvalRunner.js +648 -0
- package/dist/vendor/codali/improvement/ImprovementPolicy.d.ts +272 -0
- package/dist/vendor/codali/improvement/ImprovementPolicy.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/ImprovementPolicy.js +1450 -0
- package/dist/vendor/codali/improvement/ModelRouterCandidateBuilder.d.ts +186 -0
- package/dist/vendor/codali/improvement/ModelRouterCandidateBuilder.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/ModelRouterCandidateBuilder.js +880 -0
- package/dist/vendor/codali/improvement/OperatorInspector.d.ts +182 -0
- package/dist/vendor/codali/improvement/OperatorInspector.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/OperatorInspector.js +604 -0
- package/dist/vendor/codali/improvement/ProductionGovernance.d.ts +91 -0
- package/dist/vendor/codali/improvement/ProductionGovernance.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/ProductionGovernance.js +329 -0
- package/dist/vendor/codali/improvement/PromptSchemaToolMetadataCandidateBuilder.d.ts +152 -0
- package/dist/vendor/codali/improvement/PromptSchemaToolMetadataCandidateBuilder.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/PromptSchemaToolMetadataCandidateBuilder.js +480 -0
- package/dist/vendor/codali/improvement/PublishOrchestrator.d.ts +113 -0
- package/dist/vendor/codali/improvement/PublishOrchestrator.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/PublishOrchestrator.js +590 -0
- package/dist/vendor/codali/improvement/ReleaseOutcomeReporter.d.ts +133 -0
- package/dist/vendor/codali/improvement/ReleaseOutcomeReporter.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/ReleaseOutcomeReporter.js +501 -0
- package/dist/vendor/codali/improvement/StorageServiceImprovementClient.d.ts +71 -0
- package/dist/vendor/codali/improvement/StorageServiceImprovementClient.d.ts.map +1 -0
- package/dist/vendor/codali/improvement/StorageServiceImprovementClient.js +245 -0
- package/dist/vendor/codali/index.d.ts +43 -0
- package/dist/vendor/codali/index.d.ts.map +1 -1
- package/dist/vendor/codali/index.js +21 -0
- package/dist/vendor/codali/runtime/CodaliRuntime.d.ts +0 -2
- package/dist/vendor/codali/runtime/CodaliRuntime.d.ts.map +1 -1
- package/dist/vendor/codali/runtime/CodaliRuntime.js +0 -5
- package/dist/vendor/codali/storage/CodaliDatasetPrivacyEngine.d.ts +138 -0
- package/dist/vendor/codali/storage/CodaliDatasetPrivacyEngine.d.ts.map +1 -0
- package/dist/vendor/codali/storage/CodaliDatasetPrivacyEngine.js +417 -0
- package/dist/vendor/codali/storage/CodaliFeedbackReviewIngestion.d.ts +50 -0
- package/dist/vendor/codali/storage/CodaliFeedbackReviewIngestion.d.ts.map +1 -0
- package/dist/vendor/codali/storage/CodaliFeedbackReviewIngestion.js +141 -0
- package/dist/vendor/codali/storage/CodaliStorageContracts.d.ts +337 -0
- package/dist/vendor/codali/storage/CodaliStorageContracts.d.ts.map +1 -0
- package/dist/vendor/codali/storage/CodaliStorageContracts.js +1555 -0
- package/dist/vendor/codali/storage/DatasetExportJob.d.ts +51 -0
- package/dist/vendor/codali/storage/DatasetExportJob.d.ts.map +1 -0
- package/dist/vendor/codali/storage/DatasetExportJob.js +358 -0
- package/dist/vendor/codali/storage/DatasetReviewQueue.d.ts +98 -0
- package/dist/vendor/codali/storage/DatasetReviewQueue.d.ts.map +1 -0
- package/dist/vendor/codali/storage/DatasetReviewQueue.js +426 -0
- package/dist/vendor/codali/storage/GatewayDatasetStore.d.ts +233 -0
- package/dist/vendor/codali/storage/GatewayDatasetStore.d.ts.map +1 -0
- package/dist/vendor/codali/storage/GatewayDatasetStore.js +2070 -0
- package/package.json +3 -3
|
@@ -0,0 +1,823 @@
|
|
|
1
|
+
import { createHash } from "node:crypto";
|
|
2
|
+
import { resolveCodaliGatewayAgentTiers, } from "../gateway/AgentTierResolver.js";
|
|
3
|
+
export const CODALI_FINE_TUNE_JOB_PLANNER_SCHEMA_VERSION = "codali.improvement.fine_tune_job_planner.v1";
|
|
4
|
+
export const CODALI_FINE_TUNE_PROPOSAL_ARTIFACT = "fine-tune";
|
|
5
|
+
export const CODALI_FINE_TUNE_WORKER_ROLES = [
|
|
6
|
+
"extractor",
|
|
7
|
+
"tool_router",
|
|
8
|
+
"planner",
|
|
9
|
+
"verifier",
|
|
10
|
+
"query_expander",
|
|
11
|
+
"repair",
|
|
12
|
+
"context_refiner",
|
|
13
|
+
];
|
|
14
|
+
export const CODALI_FINE_TUNE_FINAL_SYNTHESIZER_ROLE = "final_synthesizer";
|
|
15
|
+
const DEFAULT_INVENTORY_COMMAND = {
|
|
16
|
+
command: "mcoda",
|
|
17
|
+
args: ["agent", "list", "--json", "--refresh-health"],
|
|
18
|
+
};
|
|
19
|
+
const ROLE_SPECS = {
|
|
20
|
+
extractor: {
|
|
21
|
+
resolverRole: "extractor",
|
|
22
|
+
sourceArtifactTypes: ["extractor", "extractor_sft"],
|
|
23
|
+
supportedExportKinds: ["extractor-sft"],
|
|
24
|
+
},
|
|
25
|
+
tool_router: {
|
|
26
|
+
resolverRole: "router",
|
|
27
|
+
sourceArtifactTypes: [
|
|
28
|
+
"router",
|
|
29
|
+
"tool_router",
|
|
30
|
+
"tool_router_sft",
|
|
31
|
+
"model_router",
|
|
32
|
+
"model-router",
|
|
33
|
+
],
|
|
34
|
+
supportedExportKinds: ["tool-router-sft", "model-router"],
|
|
35
|
+
},
|
|
36
|
+
planner: {
|
|
37
|
+
resolverRole: "planner",
|
|
38
|
+
sourceArtifactTypes: ["planner", "planner_sft"],
|
|
39
|
+
supportedExportKinds: ["planner-sft"],
|
|
40
|
+
},
|
|
41
|
+
verifier: {
|
|
42
|
+
resolverRole: "verifier",
|
|
43
|
+
sourceArtifactTypes: ["verifier", "verifier_sft"],
|
|
44
|
+
supportedExportKinds: ["verifier-sft"],
|
|
45
|
+
},
|
|
46
|
+
query_expander: {
|
|
47
|
+
resolverRole: "query_expander",
|
|
48
|
+
sourceArtifactTypes: [
|
|
49
|
+
"query_expander",
|
|
50
|
+
"query_expander_sft",
|
|
51
|
+
"rag_reranker",
|
|
52
|
+
"rag-reranker",
|
|
53
|
+
],
|
|
54
|
+
supportedExportKinds: ["query-expander-sft", "rag-reranker"],
|
|
55
|
+
},
|
|
56
|
+
repair: {
|
|
57
|
+
resolverRole: "repair",
|
|
58
|
+
sourceArtifactTypes: ["repair", "repair_sft"],
|
|
59
|
+
supportedExportKinds: ["repair-sft"],
|
|
60
|
+
},
|
|
61
|
+
context_refiner: {
|
|
62
|
+
resolverRole: "context_refiner",
|
|
63
|
+
sourceArtifactTypes: ["context_refiner", "context_refiner_sft"],
|
|
64
|
+
supportedExportKinds: ["context-refiner-sft"],
|
|
65
|
+
},
|
|
66
|
+
final_synthesizer: {
|
|
67
|
+
resolverRole: "final_synthesizer",
|
|
68
|
+
sourceArtifactTypes: ["final_synthesizer", "final_synthesizer_sft"],
|
|
69
|
+
supportedExportKinds: [],
|
|
70
|
+
},
|
|
71
|
+
};
|
|
72
|
+
const isRecord = (value) => Boolean(value && typeof value === "object" && !Array.isArray(value));
|
|
73
|
+
const stableJson = (value) => {
|
|
74
|
+
if (Array.isArray(value))
|
|
75
|
+
return `[${value.map(stableJson).join(",")}]`;
|
|
76
|
+
if (!isRecord(value))
|
|
77
|
+
return JSON.stringify(value);
|
|
78
|
+
return `{${Object.keys(value).sort().map((key) => `${JSON.stringify(key)}:${stableJson(value[key])}`).join(",")}}`;
|
|
79
|
+
};
|
|
80
|
+
const sha256Hex = (value) => createHash("sha256").update(stableJson(value)).digest("hex");
|
|
81
|
+
const stableId = (prefix, value) => `${prefix}-${sha256Hex(value).slice(0, 16)}`;
|
|
82
|
+
const normalizeToken = (value) => value.trim().toLowerCase().replace(/[\s-]+/g, "_");
|
|
83
|
+
const uniqueSorted = (values) => Array.from(new Set(values
|
|
84
|
+
.map((value) => value?.trim())
|
|
85
|
+
.filter((value) => Boolean(value))))
|
|
86
|
+
.sort();
|
|
87
|
+
const normalizedUniqueSorted = (values) => uniqueSorted(values.map((value) => value ? normalizeToken(value) : undefined));
|
|
88
|
+
const readString = (record, keys) => {
|
|
89
|
+
for (const key of keys) {
|
|
90
|
+
const value = record?.[key];
|
|
91
|
+
if (typeof value === "string" && value.trim())
|
|
92
|
+
return value.trim();
|
|
93
|
+
}
|
|
94
|
+
return undefined;
|
|
95
|
+
};
|
|
96
|
+
const readNumber = (record, keys) => {
|
|
97
|
+
for (const key of keys) {
|
|
98
|
+
const value = record?.[key];
|
|
99
|
+
if (typeof value === "number" && Number.isFinite(value))
|
|
100
|
+
return value;
|
|
101
|
+
if (typeof value === "string" && value.trim() && Number.isFinite(Number(value))) {
|
|
102
|
+
return Number(value);
|
|
103
|
+
}
|
|
104
|
+
}
|
|
105
|
+
return undefined;
|
|
106
|
+
};
|
|
107
|
+
const objectRefFromValue = (value) => isRecord(value) &&
|
|
108
|
+
typeof value.refId === "string" &&
|
|
109
|
+
typeof value.contentHash === "string" &&
|
|
110
|
+
typeof value.byteSize === "number" &&
|
|
111
|
+
typeof value.mimeType === "string" &&
|
|
112
|
+
typeof value.deletionGroupId === "string" &&
|
|
113
|
+
isRecord(value.privacyFlags)
|
|
114
|
+
? value
|
|
115
|
+
: undefined;
|
|
116
|
+
const refsForRow = (row) => {
|
|
117
|
+
const evidenceRefs = Array.isArray(row?.evidenceRefs)
|
|
118
|
+
? row.evidenceRefs.map(objectRefFromValue)
|
|
119
|
+
.filter((ref) => Boolean(ref))
|
|
120
|
+
: [];
|
|
121
|
+
return {
|
|
122
|
+
inputRef: objectRefFromValue(row?.inputRef),
|
|
123
|
+
outputRef: objectRefFromValue(row?.outputRef),
|
|
124
|
+
evidenceRefs,
|
|
125
|
+
};
|
|
126
|
+
};
|
|
127
|
+
const allRefsForRow = (row) => {
|
|
128
|
+
const refs = refsForRow(row);
|
|
129
|
+
return [refs.inputRef, refs.outputRef, ...refs.evidenceRefs]
|
|
130
|
+
.filter((ref) => Boolean(ref));
|
|
131
|
+
};
|
|
132
|
+
const summarizeObjectRef = (ref) => ({
|
|
133
|
+
refId: ref.refId,
|
|
134
|
+
kind: ref.kind,
|
|
135
|
+
contentHash: ref.contentHash,
|
|
136
|
+
byteSize: ref.byteSize,
|
|
137
|
+
mimeType: ref.mimeType,
|
|
138
|
+
deletionGroupId: ref.deletionGroupId,
|
|
139
|
+
privacyFlags: ref.privacyFlags,
|
|
140
|
+
...(ref.uri ? { uri: ref.uri } : {}),
|
|
141
|
+
...(ref.mediaType ? { mediaType: ref.mediaType } : {}),
|
|
142
|
+
metadataKeys: isRecord(ref.metadata) ? Object.keys(ref.metadata).sort() : [],
|
|
143
|
+
});
|
|
144
|
+
const recordIdForRow = (row) => isRecord(row) ? readString(row, ["recordId", "record_id"]) : undefined;
|
|
145
|
+
const rowsByRecordId = (rows) => {
|
|
146
|
+
const output = new Map();
|
|
147
|
+
for (const row of rows) {
|
|
148
|
+
if (!isRecord(row))
|
|
149
|
+
continue;
|
|
150
|
+
const recordId = recordIdForRow(row);
|
|
151
|
+
if (recordId)
|
|
152
|
+
output.set(recordId, row);
|
|
153
|
+
}
|
|
154
|
+
return output;
|
|
155
|
+
};
|
|
156
|
+
const metadataForRow = (row) => isRecord(row?.metadata) ? row.metadata : {};
|
|
157
|
+
const privacyForRow = (row) => isRecord(row?.privacy)
|
|
158
|
+
? row.privacy
|
|
159
|
+
: undefined;
|
|
160
|
+
const nestedRecord = (record, keys) => {
|
|
161
|
+
for (const key of keys) {
|
|
162
|
+
const value = record[key];
|
|
163
|
+
if (isRecord(value))
|
|
164
|
+
return value;
|
|
165
|
+
}
|
|
166
|
+
return {};
|
|
167
|
+
};
|
|
168
|
+
const scorecardMetadataForRow = (row) => {
|
|
169
|
+
const metadata = metadataForRow(row);
|
|
170
|
+
return nestedRecord(metadata, [
|
|
171
|
+
"scorecard",
|
|
172
|
+
"evalScorecard",
|
|
173
|
+
"eval_scorecard",
|
|
174
|
+
"qualityScorecard",
|
|
175
|
+
"quality_scorecard",
|
|
176
|
+
]);
|
|
177
|
+
};
|
|
178
|
+
const roleSpecFor = (role) => ROLE_SPECS[role];
|
|
179
|
+
export const normalizeCodaliFineTuneWorkerRole = (value) => {
|
|
180
|
+
const normalized = normalizeToken(value);
|
|
181
|
+
if (normalized === "final" || normalized === "synthesizer") {
|
|
182
|
+
return CODALI_FINE_TUNE_FINAL_SYNTHESIZER_ROLE;
|
|
183
|
+
}
|
|
184
|
+
if (normalized === CODALI_FINE_TUNE_FINAL_SYNTHESIZER_ROLE) {
|
|
185
|
+
return CODALI_FINE_TUNE_FINAL_SYNTHESIZER_ROLE;
|
|
186
|
+
}
|
|
187
|
+
return CODALI_FINE_TUNE_WORKER_ROLES.includes(normalized)
|
|
188
|
+
? normalized
|
|
189
|
+
: undefined;
|
|
190
|
+
};
|
|
191
|
+
export const fineTuneArtifactTypesForRole = (role) => [...roleSpecFor(role).sourceArtifactTypes];
|
|
192
|
+
const jobKindForExportKind = (exportKind) => exportKind === "rag-reranker" || exportKind === "model-router"
|
|
193
|
+
? "preference"
|
|
194
|
+
: "sft";
|
|
195
|
+
const supportedExportForRole = (role, exportKind) => roleSpecFor(role).supportedExportKinds.includes(exportKind);
|
|
196
|
+
const policyHashFor = (role, exportKind) => `sha256:${sha256Hex({
|
|
197
|
+
role,
|
|
198
|
+
exportKind,
|
|
199
|
+
trainingAllowedRequired: true,
|
|
200
|
+
scorecardRequired: true,
|
|
201
|
+
scorecardBypassAllowed: false,
|
|
202
|
+
finalSynthesizerAllowed: false,
|
|
203
|
+
providerAutoSubmitAllowed: false,
|
|
204
|
+
}).slice(0, 24)}`;
|
|
205
|
+
const hasTrainingPermission = (row, refs) => {
|
|
206
|
+
const privacy = privacyForRow(row);
|
|
207
|
+
return privacy?.trainingAllowed === true &&
|
|
208
|
+
refs.length > 0 &&
|
|
209
|
+
refs.every((ref) => ref.privacyFlags.trainingAllowed === true);
|
|
210
|
+
};
|
|
211
|
+
const privacySummaryForRow = (row, refs) => {
|
|
212
|
+
const privacy = privacyForRow(row);
|
|
213
|
+
return {
|
|
214
|
+
trainingAllowed: true,
|
|
215
|
+
policyTags: uniqueSorted(privacy?.policyTags ?? []),
|
|
216
|
+
containsPersonalData: privacy?.containsPersonalData === true ||
|
|
217
|
+
refs.some((ref) => ref.privacyFlags.containsPersonalData),
|
|
218
|
+
containsTenantPrivateData: refs.some((ref) => ref.privacyFlags.containsTenantPrivateData),
|
|
219
|
+
containsCustomerData: refs.some((ref) => ref.privacyFlags.containsCustomerData),
|
|
220
|
+
containsSourceCode: refs.some((ref) => ref.privacyFlags.containsSourceCode),
|
|
221
|
+
};
|
|
222
|
+
};
|
|
223
|
+
const scorecardMetricDeltas = (scorecard) => {
|
|
224
|
+
const metricNames = new Set();
|
|
225
|
+
for (const key of Object.keys(scorecard)) {
|
|
226
|
+
const normalized = key.replace(/[_-]/g, "");
|
|
227
|
+
const match = normalized.match(/^(.*?)(Before|After|Baseline|Candidate|Delta)$/i);
|
|
228
|
+
if (match?.[1]) {
|
|
229
|
+
metricNames.add(normalizeToken(match[1]));
|
|
230
|
+
}
|
|
231
|
+
}
|
|
232
|
+
const deltas = [];
|
|
233
|
+
for (const metric of [...metricNames].sort()) {
|
|
234
|
+
const before = readNumber(scorecard, [
|
|
235
|
+
`${metric}Before`,
|
|
236
|
+
`${metric}_before`,
|
|
237
|
+
`${metric}Baseline`,
|
|
238
|
+
`${metric}_baseline`,
|
|
239
|
+
]);
|
|
240
|
+
const after = readNumber(scorecard, [
|
|
241
|
+
`${metric}After`,
|
|
242
|
+
`${metric}_after`,
|
|
243
|
+
`${metric}Candidate`,
|
|
244
|
+
`${metric}_candidate`,
|
|
245
|
+
]);
|
|
246
|
+
const explicitDelta = readNumber(scorecard, [`${metric}Delta`, `${metric}_delta`]);
|
|
247
|
+
const delta = explicitDelta ?? (before !== undefined && after !== undefined ? after - before : undefined);
|
|
248
|
+
if (before !== undefined || after !== undefined || delta !== undefined) {
|
|
249
|
+
deltas.push({
|
|
250
|
+
metric,
|
|
251
|
+
...(before !== undefined ? { before } : {}),
|
|
252
|
+
...(after !== undefined ? { after } : {}),
|
|
253
|
+
...(delta !== undefined ? { delta: Number(delta.toFixed(6)) } : {}),
|
|
254
|
+
});
|
|
255
|
+
}
|
|
256
|
+
}
|
|
257
|
+
return deltas;
|
|
258
|
+
};
|
|
259
|
+
const scorecardForRow = (accepted, row) => {
|
|
260
|
+
const scorecard = scorecardMetadataForRow(row);
|
|
261
|
+
const metricKeys = Object.keys(scorecard).sort();
|
|
262
|
+
return {
|
|
263
|
+
scorecardId: stableId("fine-tune-scorecard", {
|
|
264
|
+
recordId: accepted.recordId,
|
|
265
|
+
metricKeys,
|
|
266
|
+
lineageKey: accepted.lineageKey,
|
|
267
|
+
}),
|
|
268
|
+
present: metricKeys.length > 0,
|
|
269
|
+
metricKeys,
|
|
270
|
+
metricDeltas: scorecardMetricDeltas(scorecard),
|
|
271
|
+
};
|
|
272
|
+
};
|
|
273
|
+
const tokenEstimateForRefs = (refs) => Math.max(1, Math.ceil(refs.reduce((sum, ref) => sum + ref.byteSize, 0) / 4));
|
|
274
|
+
const sourceExampleForAccepted = (accepted, row) => {
|
|
275
|
+
const refs = refsForRow(row);
|
|
276
|
+
const allRefs = [refs.inputRef, refs.outputRef, ...refs.evidenceRefs]
|
|
277
|
+
.filter((ref) => Boolean(ref));
|
|
278
|
+
if (!hasTrainingPermission(row, allRefs)) {
|
|
279
|
+
return undefined;
|
|
280
|
+
}
|
|
281
|
+
const metadata = metadataForRow(row);
|
|
282
|
+
const scorecard = scorecardForRow(accepted, row);
|
|
283
|
+
return {
|
|
284
|
+
recordId: accepted.recordId,
|
|
285
|
+
...(accepted.sourceGatewayRecordId ? {
|
|
286
|
+
sourceGatewayRecordId: accepted.sourceGatewayRecordId,
|
|
287
|
+
} : {}),
|
|
288
|
+
artifactTypes: accepted.artifactTypes,
|
|
289
|
+
preferenceSignals: accepted.preferenceSignals.map(String).sort(),
|
|
290
|
+
priorityScore: accepted.priorityScore,
|
|
291
|
+
lineageKey: accepted.lineageKey,
|
|
292
|
+
privacy: privacySummaryForRow(row, allRefs),
|
|
293
|
+
scorecard,
|
|
294
|
+
objectRefs: {
|
|
295
|
+
...(refs.inputRef ? { inputRef: summarizeObjectRef(refs.inputRef) } : {}),
|
|
296
|
+
...(refs.outputRef ? { outputRef: summarizeObjectRef(refs.outputRef) } : {}),
|
|
297
|
+
evidenceRefs: refs.evidenceRefs.map(summarizeObjectRef),
|
|
298
|
+
},
|
|
299
|
+
tokenEstimate: tokenEstimateForRefs(allRefs),
|
|
300
|
+
metadataShape: {
|
|
301
|
+
keys: Object.keys(metadata).sort(),
|
|
302
|
+
scorecardKeys: scorecard.metricKeys,
|
|
303
|
+
},
|
|
304
|
+
};
|
|
305
|
+
};
|
|
306
|
+
const buildSourceExampleSet = (inspection) => {
|
|
307
|
+
const rowMap = rowsByRecordId(inspection.primaryArtifactRows);
|
|
308
|
+
const sourceExamples = [];
|
|
309
|
+
const excludedTrainingBlockedRecordIds = [];
|
|
310
|
+
for (const accepted of inspection.curationReport.accepted) {
|
|
311
|
+
const row = rowMap.get(accepted.recordId);
|
|
312
|
+
const refs = allRefsForRow(row);
|
|
313
|
+
if (!hasTrainingPermission(row, refs)) {
|
|
314
|
+
excludedTrainingBlockedRecordIds.push(accepted.recordId);
|
|
315
|
+
continue;
|
|
316
|
+
}
|
|
317
|
+
const sourceExample = sourceExampleForAccepted(accepted, row);
|
|
318
|
+
if (sourceExample)
|
|
319
|
+
sourceExamples.push(sourceExample);
|
|
320
|
+
}
|
|
321
|
+
return {
|
|
322
|
+
sourceExamples: sourceExamples.sort((left, right) => right.priorityScore - left.priorityScore || left.recordId.localeCompare(right.recordId)),
|
|
323
|
+
excludedTrainingBlockedRecordIds: excludedTrainingBlockedRecordIds.sort(),
|
|
324
|
+
};
|
|
325
|
+
};
|
|
326
|
+
const buildTrainingManifest = (inspection, role, jobKind, sourceExamples, excludedTrainingBlockedRecordIds) => {
|
|
327
|
+
const policyHash = policyHashFor(role, inspection.manifest.exportKind);
|
|
328
|
+
const manifestId = stableId("fine-tune-manifest", {
|
|
329
|
+
exportId: inspection.exportId,
|
|
330
|
+
sourceManifestId: inspection.manifest.manifestId,
|
|
331
|
+
checksum: inspection.manifest.checksum,
|
|
332
|
+
role,
|
|
333
|
+
jobKind,
|
|
334
|
+
policyHash,
|
|
335
|
+
recordIds: sourceExamples.map((example) => example.recordId),
|
|
336
|
+
});
|
|
337
|
+
const records = sourceExamples.map((example) => {
|
|
338
|
+
const objectRefHashes = uniqueSorted([
|
|
339
|
+
example.objectRefs.inputRef?.contentHash,
|
|
340
|
+
example.objectRefs.outputRef?.contentHash,
|
|
341
|
+
...example.objectRefs.evidenceRefs.map((ref) => ref.contentHash),
|
|
342
|
+
]).map((hash) => `sha256:${hash.replace(/^sha256:/, "")}`);
|
|
343
|
+
return {
|
|
344
|
+
recordId: example.recordId,
|
|
345
|
+
...(example.sourceGatewayRecordId ? {
|
|
346
|
+
sourceGatewayRecordId: example.sourceGatewayRecordId,
|
|
347
|
+
} : {}),
|
|
348
|
+
lineageKey: example.lineageKey,
|
|
349
|
+
scorecardId: example.scorecard.scorecardId,
|
|
350
|
+
objectRefHashes,
|
|
351
|
+
tokenEstimate: example.tokenEstimate,
|
|
352
|
+
preferenceSignals: example.preferenceSignals,
|
|
353
|
+
priorityScore: example.priorityScore,
|
|
354
|
+
};
|
|
355
|
+
});
|
|
356
|
+
return {
|
|
357
|
+
manifestId,
|
|
358
|
+
jobKind,
|
|
359
|
+
role,
|
|
360
|
+
exportId: inspection.exportId,
|
|
361
|
+
sourceManifestId: inspection.manifest.manifestId,
|
|
362
|
+
sourceChecksum: inspection.manifest.checksum,
|
|
363
|
+
reproducibleFrom: {
|
|
364
|
+
exportId: inspection.exportId,
|
|
365
|
+
sourceManifestId: inspection.manifest.manifestId,
|
|
366
|
+
sourceChecksum: inspection.manifest.checksum,
|
|
367
|
+
policyHash,
|
|
368
|
+
role,
|
|
369
|
+
},
|
|
370
|
+
records,
|
|
371
|
+
rowCount: records.length,
|
|
372
|
+
totalTokenEstimate: records.reduce((sum, record) => sum + record.tokenEstimate, 0),
|
|
373
|
+
excludedTrainingBlockedRecordIds: [...excludedTrainingBlockedRecordIds],
|
|
374
|
+
};
|
|
375
|
+
};
|
|
376
|
+
const averageMetricDeltas = (sourceExamples) => {
|
|
377
|
+
const byMetric = new Map();
|
|
378
|
+
for (const example of sourceExamples) {
|
|
379
|
+
for (const delta of example.scorecard.metricDeltas) {
|
|
380
|
+
if (delta.delta === undefined)
|
|
381
|
+
continue;
|
|
382
|
+
const values = byMetric.get(delta.metric) ?? [];
|
|
383
|
+
values.push(delta.delta);
|
|
384
|
+
byMetric.set(delta.metric, values);
|
|
385
|
+
}
|
|
386
|
+
}
|
|
387
|
+
return [...byMetric.entries()].sort(([left], [right]) => left.localeCompare(right))
|
|
388
|
+
.map(([metric, values]) => ({
|
|
389
|
+
metric,
|
|
390
|
+
delta: Number((values.reduce((sum, value) => sum + value, 0) / values.length).toFixed(6)),
|
|
391
|
+
}));
|
|
392
|
+
};
|
|
393
|
+
const buildScorecardSummary = (sourceExamples) => {
|
|
394
|
+
const missingRecordIds = sourceExamples
|
|
395
|
+
.filter((example) => !example.scorecard.present)
|
|
396
|
+
.map((example) => example.recordId)
|
|
397
|
+
.sort();
|
|
398
|
+
const scorecardIds = sourceExamples
|
|
399
|
+
.filter((example) => example.scorecard.present)
|
|
400
|
+
.map((example) => example.scorecard.scorecardId)
|
|
401
|
+
.sort();
|
|
402
|
+
return {
|
|
403
|
+
required: true,
|
|
404
|
+
bypassAllowed: false,
|
|
405
|
+
examplesWithScorecards: scorecardIds.length,
|
|
406
|
+
missingRecordIds,
|
|
407
|
+
scorecardIds,
|
|
408
|
+
averageMetricDeltas: averageMetricDeltas(sourceExamples),
|
|
409
|
+
};
|
|
410
|
+
};
|
|
411
|
+
const inventorySourceFor = (input) => {
|
|
412
|
+
if (input.inventorySource)
|
|
413
|
+
return input.inventorySource;
|
|
414
|
+
if (input.inventory) {
|
|
415
|
+
return {
|
|
416
|
+
source: "provided",
|
|
417
|
+
status: "succeeded",
|
|
418
|
+
inventoryCount: input.inventory.length,
|
|
419
|
+
};
|
|
420
|
+
}
|
|
421
|
+
return {
|
|
422
|
+
source: "not_provided",
|
|
423
|
+
command: DEFAULT_INVENTORY_COMMAND.command,
|
|
424
|
+
args: [...DEFAULT_INVENTORY_COMMAND.args],
|
|
425
|
+
status: "not_run",
|
|
426
|
+
inventoryCount: 0,
|
|
427
|
+
};
|
|
428
|
+
};
|
|
429
|
+
const summarizeAssignment = (resolverRole, assignment) => ({
|
|
430
|
+
resolverRole,
|
|
431
|
+
agentSlug: assignment.candidate.slug,
|
|
432
|
+
adapter: assignment.candidate.adapter,
|
|
433
|
+
...(assignment.candidate.provider ? { provider: assignment.candidate.provider } : {}),
|
|
434
|
+
model: assignment.candidate.model,
|
|
435
|
+
source: assignment.candidate.source,
|
|
436
|
+
tier: assignment.candidate.tier,
|
|
437
|
+
runtimeHealth: {
|
|
438
|
+
status: assignment.candidate.healthStatus,
|
|
439
|
+
...(assignment.candidate.latencyMs !== undefined ? {
|
|
440
|
+
latencyMs: assignment.candidate.latencyMs,
|
|
441
|
+
} : {}),
|
|
442
|
+
},
|
|
443
|
+
capabilities: assignment.candidate.capabilities,
|
|
444
|
+
...(assignment.candidate.contextWindow !== undefined ? {
|
|
445
|
+
contextWindow: assignment.candidate.contextWindow,
|
|
446
|
+
} : {}),
|
|
447
|
+
...(assignment.candidate.costPerMillion !== undefined ? {
|
|
448
|
+
costPerMillion: assignment.candidate.costPerMillion,
|
|
449
|
+
} : {}),
|
|
450
|
+
score: Number(assignment.score.toFixed(6)),
|
|
451
|
+
reasons: [...assignment.reasons],
|
|
452
|
+
});
|
|
453
|
+
const resolveTarget = (input, role) => {
|
|
454
|
+
const spec = roleSpecFor(role);
|
|
455
|
+
const inventory = input.inventory ?? [];
|
|
456
|
+
const inventorySource = inventorySourceFor(input);
|
|
457
|
+
const refreshWarnings = input.inventoryWarnings ?? [];
|
|
458
|
+
let resolution;
|
|
459
|
+
if (inventory.length > 0) {
|
|
460
|
+
resolution = resolveCodaliGatewayAgentTiers({
|
|
461
|
+
inventory,
|
|
462
|
+
roles: [spec.resolverRole],
|
|
463
|
+
allowImageWorker: false,
|
|
464
|
+
});
|
|
465
|
+
}
|
|
466
|
+
const assignment = resolution?.assignments[spec.resolverRole];
|
|
467
|
+
return {
|
|
468
|
+
source: {
|
|
469
|
+
...inventorySource,
|
|
470
|
+
inventoryCount: inventorySource.inventoryCount ?? inventory.length,
|
|
471
|
+
},
|
|
472
|
+
status: assignment ? "resolved" : "unresolved",
|
|
473
|
+
resolverRole: spec.resolverRole,
|
|
474
|
+
...(inventory.length > 0 ? {
|
|
475
|
+
inventorySnapshotHash: `sha256:${sha256Hex(inventory).slice(0, 24)}`,
|
|
476
|
+
} : {}),
|
|
477
|
+
inventoryCount: inventory.length,
|
|
478
|
+
...(assignment ? { assignment: summarizeAssignment(spec.resolverRole, assignment) } : {}),
|
|
479
|
+
diagnostics: resolution?.diagnostics ?? [],
|
|
480
|
+
warnings: resolution?.warnings ?? [],
|
|
481
|
+
errors: resolution?.errors ?? (inventory.length === 0
|
|
482
|
+
? [{
|
|
483
|
+
code: "GATEWAY_AGENT_INVENTORY_EMPTY",
|
|
484
|
+
message: "No mcoda agent candidates were available to resolve fine-tune target.",
|
|
485
|
+
role: spec.resolverRole,
|
|
486
|
+
}]
|
|
487
|
+
: []),
|
|
488
|
+
refreshWarnings,
|
|
489
|
+
};
|
|
490
|
+
};
|
|
491
|
+
const buildEvalCommands = (role, trainingManifest, scorecardSummary) => [
|
|
492
|
+
{
|
|
493
|
+
commandId: stableId("fine-tune-eval-command", {
|
|
494
|
+
kind: "phase-27-run-tests",
|
|
495
|
+
role,
|
|
496
|
+
manifestId: trainingManifest.manifestId,
|
|
497
|
+
}),
|
|
498
|
+
command: "docdexd",
|
|
499
|
+
args: [
|
|
500
|
+
"run-tests",
|
|
501
|
+
"--repo",
|
|
502
|
+
"<repo>",
|
|
503
|
+
"--target",
|
|
504
|
+
"packages/codali/src/improvement/__tests__/FineTuneJobPlanner.test.ts",
|
|
505
|
+
],
|
|
506
|
+
dryRunSafe: true,
|
|
507
|
+
expectedScorecards: scorecardSummary.scorecardIds,
|
|
508
|
+
scorecardBypassAllowed: false,
|
|
509
|
+
},
|
|
510
|
+
{
|
|
511
|
+
commandId: stableId("fine-tune-eval-command", {
|
|
512
|
+
kind: "phase-27-pre-commit",
|
|
513
|
+
role,
|
|
514
|
+
manifestId: trainingManifest.manifestId,
|
|
515
|
+
}),
|
|
516
|
+
command: "docdexd",
|
|
517
|
+
args: [
|
|
518
|
+
"hook",
|
|
519
|
+
"pre-commit",
|
|
520
|
+
"--repo",
|
|
521
|
+
"<repo>",
|
|
522
|
+
],
|
|
523
|
+
dryRunSafe: true,
|
|
524
|
+
expectedScorecards: [],
|
|
525
|
+
scorecardBypassAllowed: false,
|
|
526
|
+
},
|
|
527
|
+
];
|
|
528
|
+
const buildEvalPlan = (role, trainingManifest, scorecardSummary, targetResolution, blockedReasons) => ({
|
|
529
|
+
planId: stableId("fine-tune-eval-plan", {
|
|
530
|
+
role,
|
|
531
|
+
manifestId: trainingManifest.manifestId,
|
|
532
|
+
scorecardIds: scorecardSummary.scorecardIds,
|
|
533
|
+
}),
|
|
534
|
+
scorecardRequired: true,
|
|
535
|
+
bypassAllowed: false,
|
|
536
|
+
scorecardIds: scorecardSummary.scorecardIds,
|
|
537
|
+
gates: [
|
|
538
|
+
{
|
|
539
|
+
gateId: stableId("fine-tune-gate", {
|
|
540
|
+
role,
|
|
541
|
+
manifestId: trainingManifest.manifestId,
|
|
542
|
+
type: "privacy",
|
|
543
|
+
}),
|
|
544
|
+
type: "privacy",
|
|
545
|
+
required: true,
|
|
546
|
+
passed: trainingManifest.rowCount > 0 &&
|
|
547
|
+
trainingManifest.excludedTrainingBlockedRecordIds.length === 0,
|
|
548
|
+
reasons: trainingManifest.excludedTrainingBlockedRecordIds.length
|
|
549
|
+
? ["training_allowed_required"]
|
|
550
|
+
: [],
|
|
551
|
+
},
|
|
552
|
+
{
|
|
553
|
+
gateId: stableId("fine-tune-gate", {
|
|
554
|
+
role,
|
|
555
|
+
manifestId: trainingManifest.manifestId,
|
|
556
|
+
type: "scorecard",
|
|
557
|
+
}),
|
|
558
|
+
type: "scorecard",
|
|
559
|
+
required: true,
|
|
560
|
+
passed: scorecardSummary.missingRecordIds.length === 0 &&
|
|
561
|
+
scorecardSummary.examplesWithScorecards === trainingManifest.rowCount &&
|
|
562
|
+
trainingManifest.rowCount > 0,
|
|
563
|
+
reasons: scorecardSummary.missingRecordIds.length
|
|
564
|
+
? ["scorecard_required"]
|
|
565
|
+
: [],
|
|
566
|
+
},
|
|
567
|
+
{
|
|
568
|
+
gateId: stableId("fine-tune-gate", {
|
|
569
|
+
role,
|
|
570
|
+
manifestId: trainingManifest.manifestId,
|
|
571
|
+
type: "target_health",
|
|
572
|
+
}),
|
|
573
|
+
type: "target_health",
|
|
574
|
+
required: true,
|
|
575
|
+
passed: targetResolution.status === "resolved",
|
|
576
|
+
reasons: targetResolution.status === "resolved" ? [] : ["target_resolution_required"],
|
|
577
|
+
},
|
|
578
|
+
{
|
|
579
|
+
gateId: stableId("fine-tune-gate", {
|
|
580
|
+
role,
|
|
581
|
+
manifestId: trainingManifest.manifestId,
|
|
582
|
+
type: "offline_eval",
|
|
583
|
+
}),
|
|
584
|
+
type: "offline_eval",
|
|
585
|
+
required: true,
|
|
586
|
+
passed: !blockedReasons.includes("scorecard_required") && trainingManifest.rowCount > 0,
|
|
587
|
+
reasons: blockedReasons.includes("scorecard_required")
|
|
588
|
+
? ["scorecard_required"]
|
|
589
|
+
: [],
|
|
590
|
+
},
|
|
591
|
+
],
|
|
592
|
+
commands: buildEvalCommands(role, trainingManifest, scorecardSummary),
|
|
593
|
+
});
|
|
594
|
+
const buildCostEstimate = (trainingManifest, targetResolution) => {
|
|
595
|
+
const costPerMillionTokens = targetResolution.assignment?.costPerMillion;
|
|
596
|
+
const estimatedProviderCostUsd = costPerMillionTokens === undefined
|
|
597
|
+
? undefined
|
|
598
|
+
: Number(((trainingManifest.totalTokenEstimate / 1000000) * costPerMillionTokens)
|
|
599
|
+
.toFixed(6));
|
|
600
|
+
return {
|
|
601
|
+
source: costPerMillionTokens === undefined ? "not_available" : "mcoda_inventory",
|
|
602
|
+
currency: "USD",
|
|
603
|
+
estimatedTrainingTokens: trainingManifest.totalTokenEstimate,
|
|
604
|
+
...(costPerMillionTokens !== undefined ? { costPerMillionTokens } : {}),
|
|
605
|
+
...(estimatedProviderCostUsd !== undefined ? { estimatedProviderCostUsd } : {}),
|
|
606
|
+
notes: costPerMillionTokens === undefined
|
|
607
|
+
? ["target cost data was not present in the mcoda inventory snapshot"]
|
|
608
|
+
: ["estimate uses mcoda inventory cost_per_million and manifest token estimate"],
|
|
609
|
+
};
|
|
610
|
+
};
|
|
611
|
+
const providerSubmissionPolicy = () => ({
|
|
612
|
+
automaticSubmission: false,
|
|
613
|
+
approvalRequired: true,
|
|
614
|
+
status: "not_submitted",
|
|
615
|
+
runnerStatus: "not_approved",
|
|
616
|
+
blocker: "provider_specific_runners_not_approved",
|
|
617
|
+
});
|
|
618
|
+
const buildRollbackPlan = (role, trainingManifest) => ({
|
|
619
|
+
planId: stableId("fine-tune-rollback-plan", {
|
|
620
|
+
role,
|
|
621
|
+
manifestId: trainingManifest.manifestId,
|
|
622
|
+
}),
|
|
623
|
+
automaticProviderJobSubmitted: false,
|
|
624
|
+
triggers: [
|
|
625
|
+
"scorecard_regression",
|
|
626
|
+
"target_health_regression",
|
|
627
|
+
"privacy_policy_revocation",
|
|
628
|
+
],
|
|
629
|
+
steps: [
|
|
630
|
+
"do_not_submit_provider_job_without_runner_approval",
|
|
631
|
+
"remove_candidate_manifest_from improvement queue",
|
|
632
|
+
"restore previous worker target assignment",
|
|
633
|
+
"rerun scorecards before any later provider submission",
|
|
634
|
+
],
|
|
635
|
+
});
|
|
636
|
+
const blockedReasonsFor = (input) => normalizedUniqueSorted([
|
|
637
|
+
input.role === CODALI_FINE_TUNE_FINAL_SYNTHESIZER_ROLE
|
|
638
|
+
? "final_synthesizer_fine_tune_disabled_by_default"
|
|
639
|
+
: undefined,
|
|
640
|
+
supportedExportForRole(input.role, input.exportKind)
|
|
641
|
+
? undefined
|
|
642
|
+
: "export_kind_not_supported_for_role",
|
|
643
|
+
input.trainingManifest.rowCount === 0 ? "no_training_allowed_examples" : undefined,
|
|
644
|
+
input.trainingManifest.excludedTrainingBlockedRecordIds.length > 0
|
|
645
|
+
? "training_allowed_required"
|
|
646
|
+
: undefined,
|
|
647
|
+
input.scorecardSummary.examplesWithScorecards === input.trainingManifest.rowCount &&
|
|
648
|
+
input.trainingManifest.rowCount > 0 &&
|
|
649
|
+
input.scorecardSummary.missingRecordIds.length === 0
|
|
650
|
+
? undefined
|
|
651
|
+
: "scorecard_required",
|
|
652
|
+
input.targetResolution.status === "resolved" ? undefined : "target_resolution_required",
|
|
653
|
+
]);
|
|
654
|
+
const buildJobSpec = (input) => {
|
|
655
|
+
const jobPlanId = stableId("fine-tune-job-plan", {
|
|
656
|
+
role: input.role,
|
|
657
|
+
sourceManifestId: input.inspection.manifest.manifestId,
|
|
658
|
+
trainingManifestId: input.trainingManifest.manifestId,
|
|
659
|
+
policyHash: input.trainingManifest.reproducibleFrom.policyHash,
|
|
660
|
+
});
|
|
661
|
+
return {
|
|
662
|
+
jobSpecId: stableId("fine-tune-job-spec", {
|
|
663
|
+
jobPlanId,
|
|
664
|
+
targetSnapshot: input.targetResolution.inventorySnapshotHash,
|
|
665
|
+
targetSlug: input.targetResolution.assignment?.agentSlug,
|
|
666
|
+
}),
|
|
667
|
+
jobPlanId,
|
|
668
|
+
status: input.blockedReasons.length ? "blocked" : "draft",
|
|
669
|
+
jobKind: input.jobKind,
|
|
670
|
+
role: input.role,
|
|
671
|
+
targetResolution: input.targetResolution,
|
|
672
|
+
trainingManifest: input.trainingManifest,
|
|
673
|
+
datasetLineage: input.inspection.manifest.lineage,
|
|
674
|
+
privacySummary: input.inspection.manifest.privacySummary,
|
|
675
|
+
scorecardSummary: input.scorecardSummary,
|
|
676
|
+
evalPlan: input.evalPlan,
|
|
677
|
+
costEstimate: input.costEstimate,
|
|
678
|
+
rollbackPlan: input.rollbackPlan,
|
|
679
|
+
providerSubmission: input.providerSubmission,
|
|
680
|
+
};
|
|
681
|
+
};
|
|
682
|
+
const buildCandidateSummary = (input) => ({
|
|
683
|
+
candidateId: stableId("fine-tune-candidate", {
|
|
684
|
+
role: input.role,
|
|
685
|
+
sourceManifestId: input.inspection.manifest.manifestId,
|
|
686
|
+
sourceRecordIds: input.sourceExamples.map((example) => example.recordId),
|
|
687
|
+
jobSpecIds: input.jobSpecs.map((spec) => spec.jobSpecId),
|
|
688
|
+
blockedReasons: input.blockedReasons,
|
|
689
|
+
}),
|
|
690
|
+
candidateKind: "fine_tune_job_spec",
|
|
691
|
+
status: input.blockedReasons.length ? "blocked" : "proposed",
|
|
692
|
+
role: input.role,
|
|
693
|
+
jobKind: input.jobKind,
|
|
694
|
+
sourceExportIds: [input.inspection.manifest.manifestId],
|
|
695
|
+
sourceRecordIds: input.sourceExamples.map((example) => example.recordId),
|
|
696
|
+
artifactIds: input.inspection.manifest.artifactRefs.map((ref) => ref.refId),
|
|
697
|
+
exampleCount: input.sourceExamples.length,
|
|
698
|
+
objectBytes: input.inspection.manifest.artifactRefs
|
|
699
|
+
.reduce((total, ref) => total + ref.byteSize, 0),
|
|
700
|
+
jobSpecIds: input.jobSpecs.map((spec) => spec.jobSpecId),
|
|
701
|
+
blockedReasons: [...input.blockedReasons],
|
|
702
|
+
scorecardRequired: true,
|
|
703
|
+
providerSubmissionEnabled: false,
|
|
704
|
+
});
|
|
705
|
+
export const buildCodaliFineTuneJobPlannerBundle = ({ inspection, role = "extractor", artifact = CODALI_FINE_TUNE_PROPOSAL_ARTIFACT, ...input }) => {
|
|
706
|
+
const jobKind = jobKindForExportKind(inspection.manifest.exportKind);
|
|
707
|
+
const { sourceExamples, excludedTrainingBlockedRecordIds } = buildSourceExampleSet(inspection);
|
|
708
|
+
const trainingManifest = buildTrainingManifest(inspection, role, jobKind, sourceExamples, excludedTrainingBlockedRecordIds);
|
|
709
|
+
const targetResolution = resolveTarget({ inspection, role, artifact, ...input }, role);
|
|
710
|
+
const scorecardSummary = buildScorecardSummary(sourceExamples);
|
|
711
|
+
const providerSubmission = providerSubmissionPolicy();
|
|
712
|
+
const initialBlockedReasons = blockedReasonsFor({
|
|
713
|
+
role,
|
|
714
|
+
exportKind: inspection.manifest.exportKind,
|
|
715
|
+
trainingManifest,
|
|
716
|
+
scorecardSummary,
|
|
717
|
+
targetResolution,
|
|
718
|
+
});
|
|
719
|
+
const evalPlan = buildEvalPlan(role, trainingManifest, scorecardSummary, targetResolution, initialBlockedReasons);
|
|
720
|
+
const costEstimate = buildCostEstimate(trainingManifest, targetResolution);
|
|
721
|
+
const rollbackPlan = buildRollbackPlan(role, trainingManifest);
|
|
722
|
+
const blockedReasons = blockedReasonsFor({
|
|
723
|
+
role,
|
|
724
|
+
exportKind: inspection.manifest.exportKind,
|
|
725
|
+
trainingManifest,
|
|
726
|
+
scorecardSummary,
|
|
727
|
+
targetResolution,
|
|
728
|
+
});
|
|
729
|
+
const jobSpecs = role === CODALI_FINE_TUNE_FINAL_SYNTHESIZER_ROLE
|
|
730
|
+
? []
|
|
731
|
+
: [
|
|
732
|
+
buildJobSpec({
|
|
733
|
+
inspection,
|
|
734
|
+
role,
|
|
735
|
+
jobKind,
|
|
736
|
+
trainingManifest,
|
|
737
|
+
targetResolution,
|
|
738
|
+
scorecardSummary,
|
|
739
|
+
evalPlan,
|
|
740
|
+
costEstimate,
|
|
741
|
+
rollbackPlan,
|
|
742
|
+
providerSubmission,
|
|
743
|
+
blockedReasons,
|
|
744
|
+
}),
|
|
745
|
+
];
|
|
746
|
+
return {
|
|
747
|
+
schemaVersion: CODALI_FINE_TUNE_JOB_PLANNER_SCHEMA_VERSION,
|
|
748
|
+
artifact,
|
|
749
|
+
source: {
|
|
750
|
+
exportId: inspection.exportId,
|
|
751
|
+
manifestId: inspection.manifest.manifestId,
|
|
752
|
+
manifestPath: inspection.manifestPath,
|
|
753
|
+
exportKind: inspection.manifest.exportKind,
|
|
754
|
+
checksum: inspection.manifest.checksum,
|
|
755
|
+
recordCount: inspection.manifest.recordCount,
|
|
756
|
+
primaryArtifactRef: summarizeObjectRef(inspection.primaryArtifact?.ref ?? inspection.manifest.artifactRefs[0]),
|
|
757
|
+
},
|
|
758
|
+
generationPolicy: {
|
|
759
|
+
deterministic: true,
|
|
760
|
+
modifiesRuntimePrompts: false,
|
|
761
|
+
modifiesRuntimeCode: false,
|
|
762
|
+
bodyPolicy: "object_refs_and_hashes_only",
|
|
763
|
+
uploadEnabled: false,
|
|
764
|
+
providerSubmissionEnabled: false,
|
|
765
|
+
scorecardRequired: true,
|
|
766
|
+
scorecardBypassAllowed: false,
|
|
767
|
+
finalSynthesizerFineTuning: false,
|
|
768
|
+
},
|
|
769
|
+
expectedShape: {
|
|
770
|
+
schemaVersion: CODALI_FINE_TUNE_JOB_PLANNER_SCHEMA_VERSION,
|
|
771
|
+
artifact,
|
|
772
|
+
requiredFields: [
|
|
773
|
+
"sourceExamples",
|
|
774
|
+
"trainingManifest",
|
|
775
|
+
"targetResolution",
|
|
776
|
+
"scorecardSummary",
|
|
777
|
+
"evalPlan",
|
|
778
|
+
"costEstimate",
|
|
779
|
+
"rollbackPlan",
|
|
780
|
+
"providerSubmission",
|
|
781
|
+
"jobSpecs",
|
|
782
|
+
],
|
|
783
|
+
jobSpecRequiredFields: [
|
|
784
|
+
"targetResolution",
|
|
785
|
+
"trainingManifest",
|
|
786
|
+
"datasetLineage",
|
|
787
|
+
"privacySummary",
|
|
788
|
+
"evalPlan",
|
|
789
|
+
"costEstimate",
|
|
790
|
+
"rollbackPlan",
|
|
791
|
+
"providerSubmission",
|
|
792
|
+
],
|
|
793
|
+
},
|
|
794
|
+
rolePolicy: {
|
|
795
|
+
role,
|
|
796
|
+
resolverRole: roleSpecFor(role).resolverRole,
|
|
797
|
+
allowedWorkerRole: role !== CODALI_FINE_TUNE_FINAL_SYNTHESIZER_ROLE,
|
|
798
|
+
supportedExportKinds: [...roleSpecFor(role).supportedExportKinds],
|
|
799
|
+
finalSynthesizerAllowed: false,
|
|
800
|
+
providerAutoSubmitAllowed: false,
|
|
801
|
+
},
|
|
802
|
+
sourceExamples,
|
|
803
|
+
trainingManifest,
|
|
804
|
+
targetResolution,
|
|
805
|
+
scorecardSummary,
|
|
806
|
+
evalPlan,
|
|
807
|
+
costEstimate,
|
|
808
|
+
rollbackPlan,
|
|
809
|
+
providerSubmission,
|
|
810
|
+
jobSpecs,
|
|
811
|
+
rejectedEvidence: inspection.curationReport.rejected,
|
|
812
|
+
candidates: [
|
|
813
|
+
buildCandidateSummary({
|
|
814
|
+
inspection,
|
|
815
|
+
role,
|
|
816
|
+
jobKind,
|
|
817
|
+
sourceExamples,
|
|
818
|
+
jobSpecs,
|
|
819
|
+
blockedReasons,
|
|
820
|
+
}),
|
|
821
|
+
],
|
|
822
|
+
};
|
|
823
|
+
};
|