@tangle-network/agent-bench 0.1.0 → 0.3.5

package/src/benchmarks/bfcl.ts

@@ -0,0 +1,346 @@
+/**
+ * Berkeley Function Calling Leaderboard adapter.
+ *
+ * Scope: deterministic function-call ground-truth categories from the official
+ * BFCL data files. This is NOT the full live BFCL leaderboard evaluator: agentic
+ * web-search/memory categories and BFCL's own model-response harness remain
+ * upstream responsibilities. The adapter loads official JSONL rows plus their
+ * `possible_answer` file and scores structured function-call artifacts against
+ * allowed function/argument values.
+ */
+
+import { readFile, stat } from 'node:fs/promises'
+import { join } from 'node:path'
+import type { OutputAdapter } from '@tangle-network/agent-runtime/loops'
+import { benchRoot } from './_harness'
+import type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'
+
+const FIXTURES = join(benchRoot, 'fixtures', 'bfcl.json')
+const DEFAULT_CATEGORY = 'BFCL_v4_simple_python'
+
+interface BfclRow {
+  id: string
+  question: unknown
+  function: unknown
+}
+
+interface BfclAnswerRow {
+  id: string
+  ground_truth: unknown
+}
+
+interface BfclAllowedCall {
+  name: string
+  arguments: Record<string, unknown[]>
+}
+
+interface BfclMeta {
+  rowId: string
+  category: string
+  functions: unknown
+  expected: BfclAllowedCall[]
+  scoring: 'bfcl-ground-truth-subset'
+}
+
+interface BfclActualCall {
+  name: string
+  arguments: Record<string, unknown>
+}
+
+const bfclDir = (): string | undefined => process.env.BFCL_DIR
+const bfclCategory = (): string => process.env.BFCL_CATEGORY ?? DEFAULT_CATEGORY
+
+export const bfclOutput: OutputAdapter<string> = {
+  parse(events) {
+    let text = ''
+    for (const ev of events) {
+      const d = (ev as { data?: Record<string, unknown> })?.data
+      const t = d?.finalText ?? d?.text ?? d?.result
+      if (typeof t === 'string' && t.length > 0) text = t
+    }
+    return text.trim()
+  },
+}
+
+async function assertPath(path: string, label: string): Promise<void> {
+  try {
+    await stat(path)
+  } catch (err) {
+    throw new Error(`BFCL: missing ${label} at ${path} (${err instanceof Error ? err.message : err})`)
+  }
+}
+
+function readJsonl(raw: string): unknown[] {
+  return raw
+    .split(/\r?\n/)
+    .map((line) => line.trim())
+    .filter((line) => line.length > 0)
+    .map((line) => JSON.parse(line) as unknown)
+}
+
+function taskFile(dir: string, category: string): string {
+  return process.env.BFCL_DATA_FILE ?? join(dir, 'bfcl_eval', 'data', `${category}.json`)
+}
+
+function answerFile(dir: string, category: string): string {
+  return process.env.BFCL_ANSWER_FILE ?? join(dir, 'bfcl_eval', 'data', 'possible_answer', `${category}.json`)
+}
+
+function questionText(question: unknown): string {
+  if (typeof question === 'string') return question
+  if (!Array.isArray(question)) return JSON.stringify(question, null, 2)
+  const turns: string[] = []
+  for (const item of question.flat(3)) {
+    if (item && typeof item === 'object') {
+      const role = typeof (item as { role?: unknown }).role === 'string' ? (item as { role: string }).role : 'user'
+      const content = (item as { content?: unknown }).content
+      if (typeof content === 'string') turns.push(`${role}: ${content}`)
+    }
+  }
+  return turns.length > 0 ? turns.join('\n') : JSON.stringify(question, null, 2)
+}
+
+function normalizeScalar(value: unknown): string {
+  if (typeof value === 'number') return Number.isInteger(value) ? String(value) : String(Number(value.toFixed(8)))
+  if (typeof value === 'string') return value.trim().toLowerCase()
+  return JSON.stringify(value)
+}
+
+function normalizeAllowed(value: unknown): unknown[] {
+  return Array.isArray(value) ? value : [value]
+}
+
+function groundTruthToCalls(value: unknown): BfclAllowedCall[] {
+  if (!Array.isArray(value)) return []
+  const out: BfclAllowedCall[] = []
+  for (const item of value) {
+    if (!item || typeof item !== 'object') continue
+    for (const [name, args] of Object.entries(item as Record<string, unknown>)) {
+      if (!args || typeof args !== 'object' || Array.isArray(args)) continue
+      const normalized: Record<string, unknown[]> = {}
+      for (const [argName, allowed] of Object.entries(args as Record<string, unknown>)) {
+        normalized[argName] = normalizeAllowed(allowed)
+      }
+      out.push({ name, arguments: normalized })
+    }
+  }
+  return out
+}
+
+function rowToTask(row: BfclRow, answer: BfclAnswerRow, category: string): BenchTask {
+  const expected = groundTruthToCalls(answer.ground_truth)
+  if (expected.length === 0) {
+    throw new Error(`BFCL ${row.id}: possible_answer has no deterministic ground_truth calls`)
+  }
+  const meta: BfclMeta = {
+    rowId: row.id,
+    category,
+    functions: row.function,
+    expected,
+    scoring: 'bfcl-ground-truth-subset',
+  }
+  return {
+    id: row.id,
+    split: category,
+    prompt: [
+      'Solve this BFCL function-calling task.',
+      'Return only JSON: {"function_calls":[{"name":"...","arguments":{...}}]}.',
+      '',
+      questionText(row.question),
+      '',
+      `Available functions: ${JSON.stringify(row.function, null, 2)}`,
+    ].join('\n'),
+    metadata: meta as unknown as Record<string, unknown>,
+  }
+}
+
+function readMeta(task: BenchTask): BfclMeta {
+  const md = task.metadata
+  if (!md || !Array.isArray(md.expected)) {
+    throw new Error(`BFCL task ${task.id} missing expected calls — loadTasks did not populate metadata`)
+  }
+  return md as unknown as BfclMeta
+}
+
+function selectTasks(tasks: BenchTask[], opts: LoadOptions): BenchTask[] {
+  let out = tasks
+  if (opts.split) out = out.filter((task) => task.split === opts.split)
+  if (opts.ids) {
+    const want = new Set(opts.ids)
+    out = out.filter((task) => want.has(task.id))
+  } else if (opts.limit !== undefined) {
+    out = out.slice(0, opts.limit)
+  }
+  if (out.length === 0) throw new Error(`BFCL: no tasks matched ${JSON.stringify(opts)}`)
+  return out
+}
+
+function buildTasks(rows: BfclRow[], answers: BfclAnswerRow[], category: string, opts: LoadOptions): BenchTask[] {
+  const byId = new Map(answers.map((answer) => [answer.id, answer]))
+  const tasks = rows.map((row) => {
+    const answer = byId.get(row.id)
+    if (!answer) throw new Error(`BFCL ${category}: missing possible_answer for ${row.id}`)
+    return rowToTask(row, answer, category)
+  })
+  return selectTasks(tasks, opts)
+}
+
+async function loadFixtures(opts: LoadOptions): Promise<BenchTask[]> {
+  const fixture = JSON.parse(await readFile(FIXTURES, 'utf8')) as { rows: BfclRow[]; answers: BfclAnswerRow[]; category: string }
+  console.warn(`[bfcl] BFCL_FIXTURES=1 — loading ${fixture.rows.length} adapter fixtures`)
+  return buildTasks(fixture.rows, fixture.answers, fixture.category, opts)
+}
+
+async function loadOfficialTasks(dir: string, opts: LoadOptions): Promise<BenchTask[]> {
+  const category = bfclCategory()
+  const dataPath = taskFile(dir, category)
+  const answersPath = answerFile(dir, category)
+  const [rawRows, rawAnswers] = await Promise.all([readFile(dataPath, 'utf8'), readFile(answersPath, 'utf8')])
+  const rows = readJsonl(rawRows) as BfclRow[]
+  const answers = readJsonl(rawAnswers) as BfclAnswerRow[]
+  return buildTasks(rows, answers, category, opts)
+}
+
+function extractJsonBlock(text: string): unknown {
+  const fences = [...text.matchAll(/```(?:json)?\s*([\s\S]*?)```/g)]
+  const raw = (fences.at(-1)?.[1] ?? text).trim()
+  try {
+    return JSON.parse(raw)
+  } catch {
+    return undefined
+  }
+}
+
+function normalizeArguments(value: unknown): Record<string, unknown> {
+  if (typeof value === 'string') {
+    try {
+      return normalizeArguments(JSON.parse(value))
+    } catch {
+      return {}
+    }
+  }
+  if (!value || typeof value !== 'object' || Array.isArray(value)) return {}
+  return value as Record<string, unknown>
+}
+
+function callFromObject(value: unknown): BfclActualCall | undefined {
+  if (!value || typeof value !== 'object') return undefined
+  const raw = value as Record<string, unknown>
+  if (raw.function && typeof raw.function === 'object') {
+    const fn = raw.function as Record<string, unknown>
+    if (typeof fn.name === 'string') return { name: fn.name, arguments: normalizeArguments(fn.arguments) }
+  }
+  const name =
+    typeof raw.name === 'string' ? raw.name
+      : typeof raw.function_name === 'string' ? raw.function_name
+        : typeof raw.tool_name === 'string' ? raw.tool_name
+          : undefined
+  if (!name) return undefined
+  return { name, arguments: normalizeArguments(raw.arguments ?? raw.args ?? raw.parameters) }
+}
+
+function extractActualCalls(artifact: string): BfclActualCall[] {
+  const parsed = extractJsonBlock(artifact)
+  if (Array.isArray(parsed)) return parsed.map(callFromObject).filter((call): call is BfclActualCall => Boolean(call))
+  if (parsed && typeof parsed === 'object') {
+    const raw = parsed as Record<string, unknown>
+    for (const key of ['function_calls', 'tool_calls', 'calls']) {
+      if (Array.isArray(raw[key])) {
+        return raw[key].map(callFromObject).filter((call): call is BfclActualCall => Boolean(call))
+      }
+    }
+    const single = callFromObject(raw)
+    if (single) return [single]
+  }
+  return []
+}
+
+function argMatches(expectedValues: unknown[], actual: unknown): boolean {
+  return expectedValues.some((expected) => normalizeScalar(expected) === normalizeScalar(actual))
+}
+
+function callMatches(expected: BfclAllowedCall, actual: BfclActualCall): boolean {
+  if (expected.name !== actual.name) return false
+  for (const [argName, allowed] of Object.entries(expected.arguments)) {
+    if (!(argName in actual.arguments)) {
+      if (allowed.some((value) => value === '' || value === null || value === undefined)) continue
+      return false
+    }
+    if (!argMatches(allowed, actual.arguments[argName])) return false
+  }
+  return true
+}
+
+function scoreCalls(task: BenchTask, artifact: string): BenchScore {
+  const meta = readMeta(task)
+  const actual = extractActualCalls(artifact)
+  const used = new Set<number>()
+  let matched = 0
+  for (const expected of meta.expected) {
+    const index = actual.findIndex((call, i) => !used.has(i) && callMatches(expected, call))
+    if (index >= 0) {
+      used.add(index)
+      matched += 1
+    }
+  }
+  const recall = meta.expected.length === 0 ? 0 : matched / meta.expected.length
+  const precision = actual.length === 0 ? 0 : matched / actual.length
+  const score = recall
+  return {
+    resolved: recall === 1 && precision === 1,
+    score,
+    detail: JSON.stringify({
+      scoring: meta.scoring,
+      category: meta.category,
+      expected: meta.expected,
+      actual,
+      precision,
+      recall,
+      fullBfclLeaderboardScore: null,
+    }),
+  }
+}
+
+export function createBfclAdapter(): BenchmarkAdapter {
+  const fixturesMode = process.env.BFCL_FIXTURES === '1'
+
+  return {
+    name: 'bfcl',
+    output: bfclOutput,
+
+    async preflight() {
+      if (fixturesMode) return
+      const dir = bfclDir()
+      if (!dir) {
+        throw new Error(
+          'BFCL_DIR is required. Fix: clone https://github.com/ShishirPatil/gorilla, set BFCL_DIR=/path/to/gorilla/berkeley-function-call-leaderboard, and optionally set BFCL_CATEGORY=BFCL_v4_simple_python.',
+        )
+      }
+      const category = bfclCategory()
+      await assertPath(taskFile(dir, category), 'BFCL task JSONL')
+      await assertPath(answerFile(dir, category), 'BFCL possible_answer JSONL')
+      await loadOfficialTasks(dir, { limit: 1 })
+    },
+
+    async loadTasks(opts: LoadOptions = {}) {
+      if (fixturesMode) return loadFixtures(opts)
+      const dir = bfclDir()
+      if (!dir) throw new Error('BFCL_DIR is required to load official BFCL rows')
+      return loadOfficialTasks(dir, opts)
+    },
+
+    async goldArtifact(task: BenchTask) {
+      const meta = readMeta(task)
+      return JSON.stringify({
+        function_calls: meta.expected.map((call) => ({
+          name: call.name,
+          arguments: Object.fromEntries(Object.entries(call.arguments).map(([key, values]) => [key, values[0]])),
+        })),
+      }, null, 2)
+    },
+
+    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {
+      return scoreCalls(task, artifact)
+    },
+  }
+}

package/src/benchmarks/crag.ts

@@ -0,0 +1,137 @@
+/**
+ * CRAG adapter (Comprehensive RAG Benchmark).
+ *
+ * Live mode expects an official or compatible CRAG JSON/JSONL export. The
+ * adapter preserves CRAG domain/type/dynamism tags in metadata and scores final
+ * answers deterministically against the provided gold answer list.
+ */
+
+import { readFile } from 'node:fs/promises'
+import { join } from 'node:path'
+import { benchRoot } from './_harness'
+import type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'
+import {
+  FINAL_ANSWER_SENTINEL,
+  allStrings,
+  answerScoreToBenchScore,
+  firstString,
+  isObject,
+  ragAnswerOutput,
+  readJsonRows,
+  scoreAnswerArtifact,
+  selectTasks,
+  stringFrom,
+} from './rag-shared'
+
+const FIXTURES = join(benchRoot, 'fixtures', 'crag.json')
+
+interface CragMeta {
+  benchmark: 'crag'
+  query: string
+  goldAnswers: string[]
+  domain: string
+  questionType: string
+  dynamism: string
+}
+
+const dataFile = (): string | undefined => process.env.CRAG_DATA_FILE
+
+function rowToTask(raw: unknown, index: number): BenchTask {
+  if (!isObject(raw)) throw new Error(`CRAG row ${index} must be an object`)
+  const query = firstString(raw, ['query', 'question', 'prompt'])
+  const goldAnswers = allStrings(raw, ['answer', 'answers', 'gold', 'gold_answer', 'expected_answer'])
+  if (!query) throw new Error(`CRAG row ${index} missing query`)
+  if (goldAnswers.length === 0) throw new Error(`CRAG row ${index} missing gold answer`)
+  const domain = stringFrom(raw.domain) ?? 'unknown'
+  const questionType = stringFrom(raw.question_type) ?? stringFrom(raw.questionType) ?? 'unknown'
+  const dynamism = stringFrom(raw.static_or_dynamic) ?? stringFrom(raw.dynamism) ?? 'unknown'
+  const id = stringFrom(raw.id) ?? stringFrom(raw.query_id) ?? `crag-${index}`
+  const meta: CragMeta = {
+    benchmark: 'crag',
+    query,
+    goldAnswers,
+    domain,
+    questionType,
+    dynamism,
+  }
+  return {
+    id,
+    split: stringFrom(raw.split) ?? domain,
+    prompt: [
+      'Answer this CRAG factual question.',
+      'Return a concise answer and do not guess when the evidence is insufficient.',
+      'End with a single final line: `FINAL ANSWER: <answer>`.',
+      '',
+      `Question: ${query}`,
+      `Domain: ${domain}`,
+      `Question type: ${questionType}`,
+      `Dynamism: ${dynamism}`,
+    ].join('\n'),
+    metadata: meta as unknown as Record<string, unknown>,
+  }
+}
+
+function readMeta(task: BenchTask): CragMeta {
+  const md = task.metadata
+  if (!md || !Array.isArray(md.goldAnswers)) {
+    throw new Error(`CRAG task ${task.id} missing metadata — loadTasks did not populate it`)
+  }
+  return md as unknown as CragMeta
+}
+
+async function loadRows(path: string): Promise<unknown[]> {
+  const rows = await readJsonRows(path)
+  if (rows.length === 0) throw new Error(`CRAG: no rows in ${path}`)
+  return rows
+}
+
+async function loadFixtures(opts: LoadOptions): Promise<BenchTask[]> {
+  const rows = JSON.parse(await readFile(FIXTURES, 'utf8')) as unknown[]
+  console.warn(`[crag] CRAG_FIXTURES=1 — loading ${rows.length} adapter fixtures`)
+  return selectTasks(rows.map(rowToTask), opts, 'CRAG')
+}
+
+export function createCragAdapter(): BenchmarkAdapter {
+  const fixturesMode = process.env.CRAG_FIXTURES === '1'
+
+  return {
+    name: 'crag',
+    output: ragAnswerOutput,
+
+    async preflight() {
+      if (fixturesMode) {
+        await readFile(FIXTURES, 'utf8')
+        return
+      }
+      const path = dataFile()
+      if (!path) {
+        throw new Error(
+          'CRAG_DATA_FILE is required. Fix: export facebookresearch/CRAG rows to JSONL and set CRAG_DATA_FILE=/path/to/crag.jsonl, or set CRAG_FIXTURES=1 for adapter plumbing.',
+        )
+      }
+      await loadRows(path)
+    },
+
+    async loadTasks(opts: LoadOptions = {}) {
+      if (fixturesMode) return loadFixtures(opts)
+      const path = dataFile()
+      if (!path) throw new Error('CRAG_DATA_FILE is required to load CRAG tasks')
+      return selectTasks((await loadRows(path)).map(rowToTask), opts, 'CRAG')
+    },
+
+    async goldArtifact(task: BenchTask) {
+      return `${FINAL_ANSWER_SENTINEL} ${readMeta(task).goldAnswers[0] ?? ''}`
+    },
+
+    async judge(task: BenchTask, artifact: string): Promise<BenchScore> {
+      const meta = readMeta(task)
+      const score = scoreAnswerArtifact(artifact, meta.goldAnswers)
+      return answerScoreToBenchScore(score, {
+        benchmark: meta.benchmark,
+        domain: meta.domain,
+        questionType: meta.questionType,
+        dynamism: meta.dynamism,
+      })
+    },
+  }
+}

package/src/benchmarks/dabstep.test.mts

@@ -0,0 +1,70 @@
+/**
+ * Offline DABStep adapter test. Official live tasks need a DABStep checkout with
+ * the released dataset.csv. Fixture mode only exercises adapter plumbing; it
+ * never scores benchmark rows without the official grade.py.
+ */
+import assert from 'node:assert/strict'
+import { mkdtemp, rm } from 'node:fs/promises'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import { test } from 'node:test'
+import { createDabstepAdapter, dabstepAnswerOutput } from './dabstep'
+
+process.env.DABSTEP_FIXTURES = '1'
+
+type Events = Parameters<typeof dabstepAnswerOutput.parse>[0]
+const stream = (text: string): Events => [{ data: { finalText: text } }] as unknown as Events
+
+test('loadTasks fixtures expose DABStep prompt and resource metadata shape', async () => {
+  const adapter = createDabstepAdapter()
+  const tasks = await adapter.loadTasks({ ids: ['1'] })
+  assert.equal(tasks.length, 1)
+  assert.equal(tasks[0].id, '1')
+  assert.match(tasks[0].prompt, /DABStep data-analysis task/)
+  const meta = tasks[0].metadata as Record<string, unknown>
+  assert.equal(meta.taskId, 1)
+  assert.equal(Array.isArray(meta.golds), true)
+})
+
+test('answer OutputAdapter extracts final fenced answer when present', () => {
+  assert.equal(dabstepAnswerOutput.parse(stream('work\n```answer\n42\n```')), '42')
+  assert.equal(dabstepAnswerOutput.parse(stream('Final Answer: 42')), 'Final Answer: 42')
+})
+
+test('goldArtifact exposes the fixture oracle without scoring it as a benchmark result', async () => {
+  const adapter = createDabstepAdapter()
+  const [task] = await adapter.loadTasks({ ids: ['1'] })
+  const gold = await adapter.goldArtifact(task)
+  assert.equal(gold, '42')
+})
+
+test('judge fails loud without an official DABSTEP_DIR/grade.py', async () => {
+  const adapter = createDabstepAdapter()
+  const [task] = await adapter.loadTasks({ ids: ['1'] })
+  delete process.env.DABSTEP_DIR
+  await assert.rejects(adapter.judge(task, '42'), /DABSTEP_DIR is required/)
+})
+
+test('preflight is fixture-safe and live mode fails loud without DABSTEP_DIR', async () => {
+  const fixtureAdapter = createDabstepAdapter()
+  await fixtureAdapter.preflight()
+  delete process.env.DABSTEP_FIXTURES
+  delete process.env.DABSTEP_DIR
+  const liveAdapter = createDabstepAdapter()
+  await assert.rejects(liveAdapter.preflight(), /DABSTEP_DIR is required/)
+  process.env.DABSTEP_FIXTURES = '1'
+})
+
+test('live preflight fails loud when the checkout is missing released dataset.csv', async () => {
+  delete process.env.DABSTEP_FIXTURES
+  const dir = await mkdtemp(join(tmpdir(), 'dabstep-missing-dataset-'))
+  process.env.DABSTEP_DIR = dir
+  try {
+    const adapter = createDabstepAdapter()
+    await assert.rejects(adapter.preflight(), /released dataset\.csv/)
+  } finally {
+    process.env.DABSTEP_FIXTURES = '1'
+    delete process.env.DABSTEP_DIR
+    await rm(dir, { recursive: true, force: true })
+  }
+})