File size: 15,613 Bytes
da8db3e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
819e7c9
 
 
 
 
 
 
 
 
 
 
da8db3e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
// @ts-nocheck
// Dump current TS-adapter outputs for the cross-repo parity verifier.
//
// The `server-only` import on every lib file at line 1 throws under Node;
// the wrapper Python verifier preloads `scripts/server_only_hook.cjs`
// via NODE_OPTIONS before invoking this script.
//
// Reads a pipeline output directory (`output/`), runs each TS adapter
// against the corresponding JSON, and writes the expected payload set
// as JSON to stdout (or a file).
//
// Usage (from `general-eval-card/`):
//   pnpm tsx scripts/dump-adapter-outputs.mts \
//     --pipeline-output /Users/jchim/projects/evaleval/eval_cards_backend_pipeline/output \
//     --out /tmp/parity-expected.json
//
// Surfaces dumped: model_cards, model_cards_lite, eval_list, eval_list_lite,
// eval_summaries (per detail), model_summaries (per family),
// aggregate_eval_summaries (per suite), matrix_eval_summaries (per suite),
// developer_summaries, developers.
import { createWriteStream, readFileSync, readdirSync, writeFileSync } from "node:fs"
import { join } from "node:path"

import * as ModelDataMod from "@/lib/model-data"
import * as HfDataMod from "@/lib/hf-data"
import * as EvalProcessingMod from "@/lib/eval-processing"
import * as BenchmarkMetadataUtilsMod from "@/lib/benchmark-metadata-utils"

// Under `tsx`, the libs are CommonJS modules; ESM `import *` lifts the
// real exports onto the synthetic `default`. Pull them off there.
const ModelData: any = (ModelDataMod as any).default ?? ModelDataMod
const HfData: any = (HfDataMod as any).default ?? HfDataMod
const EvalProcessing: any =
  (EvalProcessingMod as any).default ?? EvalProcessingMod
const BenchmarkMetadataUtils: any =
  (BenchmarkMetadataUtilsMod as any).default ?? BenchmarkMetadataUtilsMod
const candidateBenchmarkKeys = BenchmarkMetadataUtils.candidateBenchmarkKeys

const hfModelCardToEvaluationCardData = ModelData.hfModelCardToEvaluationCardData
const hfEvalEntryToListItem = ModelData.hfEvalEntryToListItem
const hfEvalDetailToSummary = ModelData.hfEvalDetailToSummary
const hfDeveloperDetailToSummary = ModelData.hfDeveloperDetailToSummary
const aggregateBenchmarkSummaries = ModelData.aggregateBenchmarkSummaries
const buildSingleMetricSuiteMatrixSummary = ModelData.buildSingleMetricSuiteMatrixSummary
const flattenModelEvaluations = HfData.flattenModelEvaluations
// Mirror the request-time normalizer in `fetchModelCardsList` /
// `fetchModelCardsListLite` (lib/hf-data.ts:822, 827): every fetch runs
// `normalizeSingleModelCardEntry` to merge setup-alias variants
// ("prompt"/"fc"/"thinking") under one variant_key. Without this the
// dump's variant_count is the raw 6-variant count from disk, but the
// user-facing API returns the post-merge count (~2-3).
const normalizeSingleModelCardEntry = HfData.normalizeSingleModelCardEntry
const createModelFamilySummary = EvalProcessing.createModelFamilySummary

function parseArgs() {
  const args: Record<string, string> = {}
  for (let i = 2; i < process.argv.length; i++) {
    const arg = process.argv[i]
    if (arg.startsWith("--")) {
      const key = arg.slice(2)
      const next = process.argv[i + 1]
      if (next && !next.startsWith("--")) {
        args[key] = next
        i++
      } else {
        args[key] = "true"
      }
    }
  }
  return args
}

function readJSON<T>(path: string): T {
  return JSON.parse(readFileSync(path, "utf-8")) as T
}

function listJSONFiles(dir: string): string[] {
  try {
    return readdirSync(dir).filter((p) => p.endsWith(".json"))
  } catch {
    return []
  }
}

interface DumpedSurface {
  surface: string
  by_id: Record<string, unknown>
}

const args = parseArgs()
const pipelineRoot = args["pipeline-output"]
const outPath = args["out"]
if (!pipelineRoot) {
  console.error("Missing --pipeline-output <dir>")
  process.exit(2)
}

const modelCards = readJSON<Array<Record<string, any>>>(
  join(pipelineRoot, "model-cards.json")
)
const modelCardsLite = readJSON<Array<Record<string, any>>>(
  join(pipelineRoot, "model-cards-lite.json")
)
const evalList = readJSON<{ evals: Array<Record<string, any>> }>(
  join(pipelineRoot, "eval-list.json")
)
const evalListLite = readJSON<{ evals: Array<Record<string, any>> }>(
  join(pipelineRoot, "eval-list-lite.json")
)
const evalsDir = join(pipelineRoot, "evals")
const modelsDir = join(pipelineRoot, "models")

// Build a sync benchmark card lookup mirror of `lib/benchmark-metadata.ts`'s
// `readPipelineBenchmarkCards` β€” used for aggregate/matrix surfaces, which
// in the request-time TS path call the async `attachBenchmarkCardToSummary`.
// The pipeline parity layer (`scripts/parity_outputs.py:build_aggregate_eval_summaries`)
// uses a sync card_map; mirror that approach so the dump runs without
// network/async machinery.
const benchmarkMetadata = readJSON<Record<string, any>>(
  join(pipelineRoot, "benchmark-metadata.json")
)
const benchmarkCardMap = new Map<string, any>()
for (const card of Object.values(benchmarkMetadata)) {
  const cardObj = card as Record<string, any>
  const name = cardObj?.benchmark_details?.name
  if (!name) continue
  for (const key of candidateBenchmarkKeys(name)) {
    if (!benchmarkCardMap.has(key)) {
      benchmarkCardMap.set(key, cardObj)
    }
  }
}

function syncGetBenchmarkCard(name: string | undefined | null): any | null {
  if (!name) return null
  for (const key of candidateBenchmarkKeys(name)) {
    const card = benchmarkCardMap.get(key)
    if (card) return card
  }
  return null
}

// Sync mirror of `attachBenchmarkCardToSummary` (lib/model-data.ts). The
// async upstream awaits `getBenchmarkCard`; here we use the sync card map
// loaded from `benchmark-metadata.json` so dump-adapter-outputs stays
// fully sync (no fetch / `server-only` boundary).
function syncAttachBenchmarkCardToSummary(summary: any): any {
  if (summary?.benchmark_card) return summary
  const candidates = [
    summary?.evaluation_name,
    summary?.composite_benchmark_name,
    summary?.composite_benchmark_key,
  ]
  for (const candidate of candidates) {
    const card = syncGetBenchmarkCard(candidate)
    if (card) return { ...summary, benchmark_card: card }
  }
  return summary
}

const dumped: DumpedSurface[] = []

dumped.push({
  surface: "model_cards",
  by_id: Object.fromEntries(
    modelCards.map((card) => {
      const normalized = normalizeSingleModelCardEntry(card as any)
      const payload = hfModelCardToEvaluationCardData(normalized as any)
      // Key by the adapter-canonical `route_id` so it matches parquet's
      // scalar `model_route_id` column.
      return [payload.route_id, payload]
    })
  ),
})

// model_cards_lite β€” same adapter pipeline as model_cards, but reads the
// `-lite.json` source. Mirrors `getModelCardsLite()` in lib/model-data.ts.
dumped.push({
  surface: "model_cards_lite",
  by_id: Object.fromEntries(
    modelCardsLite.map((card) => {
      const normalized = normalizeSingleModelCardEntry(card as any)
      const payload = hfModelCardToEvaluationCardData(normalized as any)
      return [payload.route_id, payload]
    })
  ),
})

// Mirror the request-time filter in `getEvalListData` /
// `getEvalListLiteData` (lib/model-data.ts:1251, 1291): drop entries
// whose `source_data.hf_repo` starts with `example://`. Applied here so
// the parity verifier sees the same user-facing shape the parity emitter
// produces (`scripts/parity_outputs.py:_is_example_eval_entry`).
function isExampleEntry(entry: Record<string, any>): boolean {
  const repo = entry?.source_data?.hf_repo
  return typeof repo === "string" && repo.startsWith("example://")
}

dumped.push({
  surface: "eval_list",
  by_id: Object.fromEntries(
    (evalList.evals ?? [])
      .filter((entry) => !isExampleEntry(entry))
      .map((entry) => {
        const payload = hfEvalEntryToListItem(entry as any)
        return [payload.evaluation_id, payload]
      })
  ),
})

// eval_list_lite β€” same adapter as eval_list but reads from
// `eval-list-lite.json`; mirrors `getEvalListLiteData` in lib/model-data.ts
// (which also strips example entries).
dumped.push({
  surface: "eval_list_lite",
  by_id: Object.fromEntries(
    (evalListLite.evals ?? [])
      .filter((entry) => !isExampleEntry(entry))
      .map((entry) => {
        const payload = hfEvalEntryToListItem(entry as any)
        return [payload.evaluation_id, payload]
      })
  ),
})

// Sort eval-detail filenames by codepoint so iteration order matches
// Python's default string comparison on the parity side (which sorts
// `parity_outputs.build_aggregate_eval_summaries` inputs by
// `eval_summary_id`). `localeCompare` treats `_` as collation-ignorable
// at primary level, putting `foo_25.json` before `foo.json` β€” Python's
// codepoint sort orders them the other way. Plain `.sort()` (no
// compareFn) does the same codepoint comparison Python does.
const evalDetails: Array<Record<string, any>> = listJSONFiles(evalsDir)
  .slice()
  .sort()
  .map((file) => readJSON<Record<string, any>>(join(evalsDir, file)))

dumped.push({
  surface: "eval_summaries",
  by_id: Object.fromEntries(
    evalDetails.map((detail) => [
      detail.eval_summary_id,
      hfEvalDetailToSummary(detail as any),
    ])
  ),
})

// aggregate_eval_summaries β€” port of the `aggregate__<suite_key>` branch in
// `getEvalSummaryById` (lib/model-data.ts). The TS path runs each sub-eval
// through `hfEvalDetailToSummary`, attaches a benchmark card, then calls
// `aggregateBenchmarkSummaries(summaries, suiteKey)`. We mirror parity
// emitter `build_aggregate_eval_summaries` (parity_outputs.py:212-258):
//   - group by `benchmark_family_key || benchmark_parent_key`
//   - skip groups with fewer than 2 distinct sub-evals
// Keyed by `payload.evaluation_id` (= `aggregate__<suite_key>`).
{
  const aggregateGroups = new Map<string, Record<string, any>[]>()
  for (const detail of evalDetails) {
    const suiteKey =
      detail.benchmark_family_key || detail.benchmark_parent_key
    if (!suiteKey) continue
    if (!detail.eval_summary_id) continue
    const list = aggregateGroups.get(String(suiteKey)) ?? []
    list.push(detail)
    aggregateGroups.set(String(suiteKey), list)
  }

  const aggregateById: Record<string, any> = {}
  for (const [suiteKey, details] of aggregateGroups.entries()) {
    // De-dupe by eval_summary_id (parity also dedupes β€” first-write-wins).
    const seenIds = new Set<string>()
    const uniqueDetails: Record<string, any>[] = []
    for (const detail of details) {
      const id = detail.eval_summary_id
      if (seenIds.has(id)) continue
      seenIds.add(id)
      uniqueDetails.push(detail)
    }
    if (uniqueDetails.length < 2) continue

    const summaries = uniqueDetails.map((detail) => {
      const summary = hfEvalDetailToSummary(detail as any)
      return syncAttachBenchmarkCardToSummary(summary)
    })

    const aggregated = aggregateBenchmarkSummaries(summaries as any, suiteKey)
    if (!aggregated) continue
    aggregateById[aggregated.evaluation_id] = aggregated
  }

  dumped.push({
    surface: "aggregate_eval_summaries",
    by_id: aggregateById,
  })
}

// matrix_eval_summaries β€” port of the `matrix__<suite_key>` branch in
// `getEvalSummaryById`. Mirrors parity emitter `build_matrix_eval_summaries`
// (parity_outputs.py:261-281):
//   - skip details where `is_summary_score` is true
//   - group by `benchmark_family_key || benchmark_parent_key`
//   - call `buildSingleMetricSuiteMatrixSummary(details, suiteKey)`
// Keyed by `payload.evaluation_id` (= `matrix__<suite_key>`).
{
  const matrixGroups = new Map<string, Record<string, any>[]>()
  for (const detail of evalDetails) {
    if (detail.is_summary_score) continue
    const suiteKey =
      detail.benchmark_family_key || detail.benchmark_parent_key
    if (!suiteKey) continue
    const list = matrixGroups.get(String(suiteKey)) ?? []
    list.push(detail)
    matrixGroups.set(String(suiteKey), list)
  }

  const matrixById: Record<string, any> = {}
  for (const [suiteKey, details] of matrixGroups.entries()) {
    const result = buildSingleMetricSuiteMatrixSummary(details as any, suiteKey)
    if (!result) continue
    const attached = syncAttachBenchmarkCardToSummary(result)
    matrixById[attached.evaluation_id] = attached
  }

  dumped.push({
    surface: "matrix_eval_summaries",
    by_id: matrixById,
  })
}

const modelDetails: Array<Record<string, any>> = listJSONFiles(modelsDir).map((file) =>
  readJSON<Record<string, any>>(join(modelsDir, file))
)

dumped.push({
  surface: "model_summaries",
  by_id: Object.fromEntries(
    modelDetails.flatMap((detail) => {
      try {
        const evaluations = flattenModelEvaluations(detail as any)
        if (evaluations.length === 0) return []
        const payload = createModelFamilySummary(evaluations as any)
        return [[payload.model_route_id, payload]]
      } catch (error) {
        // The TS guard `assertSourceMetadata` throws when source_metadata
        // is missing on any model_result; surface as a parity-comparable
        // sentinel keyed by the canonical route id (lookup against the
        // canonical model_family_id avoids drift when the input file's
        // route_id was generated pre-canonicalization).
        const fallbackRoute = (detail.model_family_id ?? "").replace(/\//g, "__")
        return [[fallbackRoute || detail.model_route_id, { _adapter_error: String(error) }]]
      }
    })
  ),
})

// Developer surfaces β€” port of getDeveloperList / getDeveloperSummaryById.
// Both run hfDeveloperDetailToSummary against pipeline `developers/*.json`;
// the list endpoint then strips `models[]` to keep the index lightweight.
const developersDir = join(pipelineRoot, "developers")
const developerDetails: Array<Record<string, any>> = listJSONFiles(developersDir)
  .map((file) => readJSON<Record<string, any>>(join(developersDir, file)))
  .filter((detail) => detail && detail.developer && Array.isArray(detail.models))

const developerSummaries = developerDetails.map((detail) => hfDeveloperDetailToSummary(detail as any))

dumped.push({
  surface: "developer_summaries",
  by_id: Object.fromEntries(developerSummaries.map((s) => [s.route_id, s])),
})

dumped.push({
  surface: "developers",
  by_id: Object.fromEntries(
    developerSummaries.map((s) => {
      // eslint-disable-next-line @typescript-eslint/no-unused-vars
      const { models, ...listEntry } = s
      return [listEntry.route_id, listEntry]
    })
  ),
})

// Stream surfaces individually β€” `JSON.stringify` of the full dump can
// exceed Node's max string length on production-scale corpora (~5.8k
// model_summaries with nested `evaluations_by_category`).
function streamSurface(handle: NodeJS.WritableStream, surface: DumpedSurface, isFirst: boolean) {
  if (!isFirst) handle.write(",")
  handle.write(`{"surface":${JSON.stringify(surface.surface)},"by_id":{`)
  let first = true
  for (const [key, value] of Object.entries(surface.by_id)) {
    if (!first) handle.write(",")
    first = false
    handle.write(JSON.stringify(key))
    handle.write(":")
    handle.write(JSON.stringify(value))
  }
  handle.write("}}")
}

async function emit(): Promise<void> {
  if (outPath) {
    const stream = createWriteStream(outPath)
    stream.write("[")
    dumped.forEach((surface, idx) => streamSurface(stream as any, surface, idx === 0))
    stream.write("]")
    await new Promise<void>((resolve) => {
      stream.end(() => resolve())
    })
    console.log(`Wrote ${dumped.length} surfaces to ${outPath}`)
  } else {
    dumped.forEach((surface, idx) =>
      streamSurface(process.stdout, surface, idx === 0)
    )
  }
}

await emit()