microsoft/hve-core
Publicmirrored from https://github.com/microsoft/hve-coreAvailable
scripts/evals/Modules/VallyRunner.psm1
486lines · modecode
| 1 | # Copyright (c) Microsoft Corporation. |
| 2 | # SPDX-License-Identifier: MIT |
| 3 | |
| 4 | # VallyRunner.psm1 |
| 5 | # |
| 6 | # Purpose: Spawn `vally eval` for a single spec, locate the timestamped run |
| 7 | # directory vally writes under --output-dir, and aggregate the |
| 8 | # resulting results.jsonl into pass/fail counts suitable for the |
| 9 | # PR-time eval-summary report. |
| 10 | # Author: HVE Core Team |
| 11 | |
| 12 | #Requires -Version 7.0 |
| 13 | |
| 14 | Set-StrictMode -Version Latest |
| 15 | |
| 16 | function Resolve-VallyRunDir { |
| 17 | <# |
| 18 | .SYNOPSIS |
| 19 | Returns the most recently written subdirectory of an `--output-dir`. |
| 20 | |
| 21 | .DESCRIPTION |
| 22 | `vally eval` writes each invocation under a timestamped subdirectory of |
| 23 | the directory passed to `--output-dir`. Callers need the latest such |
| 24 | directory to locate `results.jsonl`. |
| 25 | |
| 26 | .PARAMETER OutputDir |
| 27 | Directory that was passed to `vally eval --output-dir`. |
| 28 | |
| 29 | .OUTPUTS |
| 30 | [string] Full path to the newest subdirectory, or $null when none exists. |
| 31 | #> |
| 32 | [CmdletBinding()] |
| 33 | [OutputType([string])] |
| 34 | param( |
| 35 | [Parameter(Mandatory = $true)] |
| 36 | [string]$OutputDir |
| 37 | ) |
| 38 | |
| 39 | if (-not (Test-Path -LiteralPath $OutputDir -PathType Container)) { return $null } |
| 40 | |
| 41 | $latest = Get-ChildItem -LiteralPath $OutputDir -Directory -ErrorAction SilentlyContinue | |
| 42 | Sort-Object LastWriteTime -Descending | |
| 43 | Select-Object -First 1 |
| 44 | |
| 45 | if (-not $latest) { return $null } |
| 46 | return $latest.FullName |
| 47 | } |
| 48 | |
| 49 | function Read-VallyResultsJsonl { |
| 50 | <# |
| 51 | .SYNOPSIS |
| 52 | Aggregates trial outcomes from a vally `results.jsonl` file. |
| 53 | |
| 54 | .DESCRIPTION |
| 55 | Reads the `results.jsonl` written by `vally eval` (located under the run |
| 56 | directory returned by `Resolve-VallyRunDir`) and tallies passing/failing |
| 57 | trials plus aggregate wall time. Malformed lines are skipped rather than |
| 58 | thrown so a partial run still yields counts. |
| 59 | |
| 60 | .PARAMETER RunDir |
| 61 | Directory returned by `Resolve-VallyRunDir`. |
| 62 | |
| 63 | .OUTPUTS |
| 64 | [hashtable] `@{ assertionsPassed; assertionsFailed; durationMs; trials; resultsPath; perStimulus }`. |
| 65 | `perStimulus` is an ordered map keyed by stimulus name with `@{ assertionsPassed; assertionsFailed; durationMs; trials }`. |
| 66 | #> |
| 67 | [CmdletBinding()] |
| 68 | [OutputType([hashtable])] |
| 69 | param( |
| 70 | [Parameter(Mandatory = $true)] |
| 71 | [AllowNull()] |
| 72 | [AllowEmptyString()] |
| 73 | [string]$RunDir |
| 74 | ) |
| 75 | |
| 76 | $empty = @{ |
| 77 | assertionsPassed = 0 |
| 78 | assertionsFailed = 0 |
| 79 | durationMs = 0 |
| 80 | trials = 0 |
| 81 | resultsPath = $null |
| 82 | perStimulus = [ordered]@{} |
| 83 | } |
| 84 | |
| 85 | if ([string]::IsNullOrWhiteSpace($RunDir) -or -not (Test-Path -LiteralPath $RunDir -PathType Container)) { |
| 86 | return $empty |
| 87 | } |
| 88 | |
| 89 | $jsonl = Get-ChildItem -LiteralPath $RunDir -Filter 'results.jsonl' -Recurse -File -ErrorAction SilentlyContinue | |
| 90 | Select-Object -First 1 |
| 91 | if (-not $jsonl) { return $empty } |
| 92 | |
| 93 | $passed = 0 |
| 94 | $failed = 0 |
| 95 | $durationMs = 0 |
| 96 | $trials = 0 |
| 97 | $perStimulus = [ordered]@{} |
| 98 | |
| 99 | foreach ($line in Get-Content -LiteralPath $jsonl.FullName -Encoding utf8) { |
| 100 | if ([string]::IsNullOrWhiteSpace($line)) { continue } |
| 101 | try { |
| 102 | $obj = $line | ConvertFrom-Json -Depth 100 |
| 103 | } |
| 104 | catch { |
| 105 | continue |
| 106 | } |
| 107 | |
| 108 | $trials++ |
| 109 | |
| 110 | $trialPassed = $false |
| 111 | if ($obj.PSObject.Properties['gradeResult'] -and $obj.gradeResult -and |
| 112 | $obj.gradeResult.PSObject.Properties['passed'] -and $null -ne $obj.gradeResult.passed) { |
| 113 | $trialPassed = [bool]$obj.gradeResult.passed |
| 114 | } |
| 115 | if ($trialPassed) { $passed++ } else { $failed++ } |
| 116 | |
| 117 | $trialWallMs = 0 |
| 118 | if ($obj.PSObject.Properties['trajectory'] -and $obj.trajectory -and |
| 119 | $obj.trajectory.PSObject.Properties['metrics'] -and $obj.trajectory.metrics -and |
| 120 | $obj.trajectory.metrics.PSObject.Properties['wallTimeMs'] -and |
| 121 | $null -ne $obj.trajectory.metrics.wallTimeMs) { |
| 122 | $trialWallMs = [int]$obj.trajectory.metrics.wallTimeMs |
| 123 | $durationMs += $trialWallMs |
| 124 | } |
| 125 | |
| 126 | $stimulusName = $null |
| 127 | if ($obj.PSObject.Properties['trajectory'] -and $obj.trajectory -and |
| 128 | $obj.trajectory.PSObject.Properties['stimulus'] -and $obj.trajectory.stimulus -and |
| 129 | $obj.trajectory.stimulus.PSObject.Properties['name'] -and |
| 130 | -not [string]::IsNullOrWhiteSpace([string]$obj.trajectory.stimulus.name)) { |
| 131 | $stimulusName = [string]$obj.trajectory.stimulus.name |
| 132 | } |
| 133 | |
| 134 | if ($stimulusName) { |
| 135 | if (-not $perStimulus.Contains($stimulusName)) { |
| 136 | $perStimulus[$stimulusName] = @{ |
| 137 | assertionsPassed = 0 |
| 138 | assertionsFailed = 0 |
| 139 | durationMs = 0 |
| 140 | trials = 0 |
| 141 | } |
| 142 | } |
| 143 | $bucket = $perStimulus[$stimulusName] |
| 144 | $bucket.trials++ |
| 145 | if ($trialPassed) { $bucket.assertionsPassed++ } else { $bucket.assertionsFailed++ } |
| 146 | $bucket.durationMs += $trialWallMs |
| 147 | } |
| 148 | } |
| 149 | |
| 150 | return @{ |
| 151 | assertionsPassed = $passed |
| 152 | assertionsFailed = $failed |
| 153 | durationMs = $durationMs |
| 154 | trials = $trials |
| 155 | resultsPath = $jsonl.FullName |
| 156 | perStimulus = $perStimulus |
| 157 | } |
| 158 | } |
| 159 | |
| 160 | function Invoke-VallySpec { |
| 161 | <# |
| 162 | .SYNOPSIS |
| 163 | Runs `vally eval` for a single spec and returns aggregated outcomes. |
| 164 | |
| 165 | .DESCRIPTION |
| 166 | Invokes the configured vally executable with `eval --eval-spec --model |
| 167 | --output-dir`, captures stdout/stderr (optionally tee'd to a log file), |
| 168 | resolves the timestamped run directory under `OutputDir`, and aggregates |
| 169 | the `results.jsonl` via `Read-VallyResultsJsonl`. |
| 170 | |
| 171 | .PARAMETER SpecPath |
| 172 | Path to the eval spec YAML file. |
| 173 | |
| 174 | .PARAMETER OutputDir |
| 175 | Directory passed to `vally eval --output-dir`. Created if it does not exist. |
| 176 | |
| 177 | .PARAMETER Model |
| 178 | Model passed to `vally eval --model`. |
| 179 | |
| 180 | .PARAMETER VallyCommand |
| 181 | Path or name of the vally executable. Defaults to `vally`. Tests override |
| 182 | this with the stub fixture path. |
| 183 | |
| 184 | .PARAMETER LogPath |
| 185 | Optional path to tee stdout/stderr to a log file. |
| 186 | |
| 187 | .OUTPUTS |
| 188 | [hashtable] `@{ specPath; exitCode; runDir; assertionsPassed; assertionsFailed; durationMs; trials; resultsPath; perStimulus }`. |
| 189 | #> |
| 190 | [CmdletBinding()] |
| 191 | [OutputType([hashtable])] |
| 192 | param( |
| 193 | [Parameter(Mandatory = $true)][string]$SpecPath, |
| 194 | [Parameter(Mandatory = $true)][string]$OutputDir, |
| 195 | [Parameter(Mandatory = $true)][string]$Model, |
| 196 | [string]$VallyCommand = 'vally', |
| 197 | [string]$LogPath |
| 198 | ) |
| 199 | |
| 200 | if (-not (Test-Path -LiteralPath $OutputDir)) { |
| 201 | New-Item -ItemType Directory -Path $OutputDir -Force | Out-Null |
| 202 | } |
| 203 | |
| 204 | $vallyArgs = @( |
| 205 | 'eval' |
| 206 | '--eval-spec', $SpecPath |
| 207 | '--model', $Model |
| 208 | '--output-dir', $OutputDir |
| 209 | ) |
| 210 | |
| 211 | $sw = [System.Diagnostics.Stopwatch]::StartNew() |
| 212 | $prev = [Console]::OutputEncoding |
| 213 | $exitCode = 0 |
| 214 | try { |
| 215 | [Console]::OutputEncoding = [System.Text.Encoding]::UTF8 |
| 216 | $raw = & $VallyCommand @vallyArgs 2>&1 |
| 217 | $exitCode = $LASTEXITCODE |
| 218 | } |
| 219 | finally { |
| 220 | [Console]::OutputEncoding = $prev |
| 221 | $sw.Stop() |
| 222 | } |
| 223 | |
| 224 | $lines = @($raw | ForEach-Object { $_.ToString() }) |
| 225 | foreach ($line in $lines) { Write-Host $line } |
| 226 | |
| 227 | if ($LogPath) { |
| 228 | $dir = Split-Path -Parent $LogPath |
| 229 | if ($dir -and -not (Test-Path -LiteralPath $dir)) { |
| 230 | New-Item -ItemType Directory -Path $dir -Force | Out-Null |
| 231 | } |
| 232 | Set-Content -LiteralPath $LogPath -Value $lines -Encoding utf8NoBOM |
| 233 | } |
| 234 | |
| 235 | $runDir = Resolve-VallyRunDir -OutputDir $OutputDir |
| 236 | $aggregate = Read-VallyResultsJsonl -RunDir $runDir |
| 237 | |
| 238 | $durationMs = if ($aggregate.durationMs -gt 0) { |
| 239 | [int]$aggregate.durationMs |
| 240 | } |
| 241 | else { |
| 242 | [int]$sw.ElapsedMilliseconds |
| 243 | } |
| 244 | |
| 245 | return @{ |
| 246 | specPath = $SpecPath |
| 247 | exitCode = $exitCode |
| 248 | runDir = $runDir |
| 249 | assertionsPassed = $aggregate.assertionsPassed |
| 250 | assertionsFailed = $aggregate.assertionsFailed |
| 251 | durationMs = $durationMs |
| 252 | trials = $aggregate.trials |
| 253 | resultsPath = $aggregate.resultsPath |
| 254 | perStimulus = $aggregate.perStimulus |
| 255 | } |
| 256 | } |
| 257 | |
| 258 | function Test-SpecInputModeration { |
| 259 | <# |
| 260 | .SYNOPSIS |
| 261 | Moderates all stimulus prompts in an eval spec before execution. |
| 262 | |
| 263 | .DESCRIPTION |
| 264 | Parses the eval spec YAML, extracts all stimulus.prompt fields, sends them |
| 265 | through Invoke-ContentModeration.ps1, and returns a moderation result that |
| 266 | indicates whether the spec should be skipped due to flagged input. |
| 267 | |
| 268 | .PARAMETER SpecPath |
| 269 | Path to the eval spec YAML file. |
| 270 | |
| 271 | .PARAMETER ArtifactId |
| 272 | Artifact identifier for scope tagging (e.g., "agent-name"). |
| 273 | |
| 274 | .PARAMETER ModerationScript |
| 275 | Path to Invoke-ContentModeration.ps1. Defaults to scripts/evals/Invoke-ContentModeration.ps1. |
| 276 | |
| 277 | .PARAMETER Threshold |
| 278 | Toxicity threshold (0.0-1.0). Defaults to 0.5. |
| 279 | |
| 280 | .PARAMETER RepoRoot |
| 281 | Repository root. Defaults to git root. |
| 282 | |
| 283 | .OUTPUTS |
| 284 | [hashtable] @{ flagged = $bool; flaggedCount = $int; outputPath = $string } |
| 285 | #> |
| 286 | [CmdletBinding()] |
| 287 | [OutputType([hashtable])] |
| 288 | param( |
| 289 | [Parameter(Mandatory = $true)][string]$SpecPath, |
| 290 | [Parameter(Mandatory = $true)][string]$ArtifactId, |
| 291 | [string]$ModerationScript, |
| 292 | [double]$Threshold = 0.5, |
| 293 | [string]$RepoRoot |
| 294 | ) |
| 295 | |
| 296 | if (-not $RepoRoot) { |
| 297 | $RepoRoot = git rev-parse --show-toplevel 2>$null |
| 298 | if (-not $RepoRoot) { $RepoRoot = Join-Path $PSScriptRoot '../../..' } |
| 299 | } |
| 300 | if (-not $ModerationScript) { |
| 301 | $ModerationScript = Join-Path $RepoRoot 'scripts/evals/Invoke-ContentModeration.ps1' |
| 302 | } |
| 303 | |
| 304 | if (-not (Test-Path -LiteralPath $SpecPath -PathType Leaf)) { |
| 305 | Write-Warning "Spec file not found: $SpecPath" |
| 306 | return @{ flagged = $false; flaggedCount = 0; outputPath = $null } |
| 307 | } |
| 308 | |
| 309 | $specContent = Get-Content -LiteralPath $SpecPath -Raw -Encoding utf8 |
| 310 | try { |
| 311 | $spec = $specContent | ConvertFrom-Yaml |
| 312 | } |
| 313 | catch { |
| 314 | Write-Warning "Failed to parse spec YAML: $SpecPath" |
| 315 | return @{ flagged = $false; flaggedCount = 0; outputPath = $null } |
| 316 | } |
| 317 | |
| 318 | $records = @() |
| 319 | $index = 0 |
| 320 | if ($spec.PSObject.Properties['stimuli'] -and $spec.stimuli) { |
| 321 | foreach ($stimulus in $spec.stimuli) { |
| 322 | if ($stimulus.PSObject.Properties['prompt'] -and $stimulus.prompt) { |
| 323 | $records += @{ |
| 324 | id = "input-$ArtifactId-$index" |
| 325 | text = [string]$stimulus.prompt |
| 326 | } |
| 327 | $index++ |
| 328 | } |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | if ($records.Count -eq 0) { |
| 333 | Write-Verbose "No stimulus prompts to moderate in $SpecPath" |
| 334 | return @{ flagged = $false; flaggedCount = 0; outputPath = $null } |
| 335 | } |
| 336 | |
| 337 | $scope = "input-$ArtifactId" |
| 338 | $outFile = Join-Path $RepoRoot "logs/moderation-$scope.json" |
| 339 | |
| 340 | Write-Verbose "Moderating $($records.Count) stimulus prompts for artifact: $ArtifactId" |
| 341 | try { |
| 342 | & $ModerationScript -Records $records -Scope $scope -Threshold $Threshold -OutFile $outFile -ErrorAction Stop |
| 343 | $moderationExitCode = $LASTEXITCODE |
| 344 | } |
| 345 | catch { |
| 346 | Write-Warning "Content moderation script failed: $_" |
| 347 | return @{ flagged = $true; flaggedCount = $records.Count; outputPath = $outFile } |
| 348 | } |
| 349 | |
| 350 | $flagged = $moderationExitCode -ne 0 |
| 351 | $flaggedCount = 0 |
| 352 | if (Test-Path -LiteralPath $outFile) { |
| 353 | $output = Get-Content -LiteralPath $outFile -Raw | ConvertFrom-Json |
| 354 | $flaggedCount = [int]$output.summary.flaggedCount |
| 355 | } |
| 356 | |
| 357 | return @{ |
| 358 | flagged = $flagged |
| 359 | flaggedCount = $flaggedCount |
| 360 | outputPath = $outFile |
| 361 | } |
| 362 | } |
| 363 | |
| 364 | function Test-SpecOutputModeration { |
| 365 | <# |
| 366 | .SYNOPSIS |
| 367 | Moderates model outputs from a vally eval results.jsonl file. |
| 368 | |
| 369 | .DESCRIPTION |
| 370 | Reads the results.jsonl from a vally run directory, extracts all trajectory |
| 371 | model outputs, sends them through Invoke-ContentModeration.ps1, and returns |
| 372 | a moderation result indicating whether the spec outputs should be flagged. |
| 373 | |
| 374 | .PARAMETER RunDir |
| 375 | Vally run directory (timestamped subdirectory under --output-dir). |
| 376 | |
| 377 | .PARAMETER ArtifactId |
| 378 | Artifact identifier for scope tagging. |
| 379 | |
| 380 | .PARAMETER ModerationScript |
| 381 | Path to Invoke-ContentModeration.ps1. |
| 382 | |
| 383 | .PARAMETER Threshold |
| 384 | Toxicity threshold (0.0-1.0). Defaults to 0.5. |
| 385 | |
| 386 | .PARAMETER RepoRoot |
| 387 | Repository root. |
| 388 | |
| 389 | .OUTPUTS |
| 390 | [hashtable] @{ flagged = $bool; flaggedCount = $int; outputPath = $string } |
| 391 | #> |
| 392 | [CmdletBinding()] |
| 393 | [OutputType([hashtable])] |
| 394 | param( |
| 395 | [Parameter(Mandatory = $true)][string]$RunDir, |
| 396 | [Parameter(Mandatory = $true)][string]$ArtifactId, |
| 397 | [string]$ModerationScript, |
| 398 | [double]$Threshold = 0.5, |
| 399 | [string]$RepoRoot |
| 400 | ) |
| 401 | |
| 402 | if (-not $RepoRoot) { |
| 403 | $RepoRoot = git rev-parse --show-toplevel 2>$null |
| 404 | if (-not $RepoRoot) { $RepoRoot = Join-Path $PSScriptRoot '../../..' } |
| 405 | } |
| 406 | if (-not $ModerationScript) { |
| 407 | $ModerationScript = Join-Path $RepoRoot 'scripts/evals/Invoke-ContentModeration.ps1' |
| 408 | } |
| 409 | |
| 410 | if ([string]::IsNullOrWhiteSpace($RunDir) -or -not (Test-Path -LiteralPath $RunDir -PathType Container)) { |
| 411 | Write-Warning "Run directory not found: $RunDir" |
| 412 | return @{ flagged = $false; flaggedCount = 0; outputPath = $null } |
| 413 | } |
| 414 | |
| 415 | $jsonl = Get-ChildItem -LiteralPath $RunDir -Filter 'results.jsonl' -Recurse -File -ErrorAction SilentlyContinue | |
| 416 | Select-Object -First 1 |
| 417 | if (-not $jsonl) { |
| 418 | Write-Warning "results.jsonl not found in $RunDir" |
| 419 | return @{ flagged = $false; flaggedCount = 0; outputPath = $null } |
| 420 | } |
| 421 | |
| 422 | $records = @() |
| 423 | $index = 0 |
| 424 | foreach ($line in Get-Content -LiteralPath $jsonl.FullName -Encoding utf8) { |
| 425 | if ([string]::IsNullOrWhiteSpace($line)) { continue } |
| 426 | try { |
| 427 | $obj = $line | ConvertFrom-Json -Depth 100 |
| 428 | } |
| 429 | catch { |
| 430 | continue |
| 431 | } |
| 432 | |
| 433 | $outputText = $null |
| 434 | if ($obj.PSObject.Properties['trajectory'] -and $obj.trajectory -and |
| 435 | $obj.trajectory.PSObject.Properties['output'] -and $obj.trajectory.output) { |
| 436 | $outputText = [string]$obj.trajectory.output |
| 437 | } |
| 438 | |
| 439 | if ($outputText) { |
| 440 | $records += @{ |
| 441 | id = "output-$ArtifactId-$index" |
| 442 | text = $outputText |
| 443 | } |
| 444 | $index++ |
| 445 | } |
| 446 | } |
| 447 | |
| 448 | if ($records.Count -eq 0) { |
| 449 | Write-Verbose "No model outputs to moderate from $($jsonl.FullName)" |
| 450 | return @{ flagged = $false; flaggedCount = 0; outputPath = $null } |
| 451 | } |
| 452 | |
| 453 | $scope = "output-$ArtifactId" |
| 454 | $outFile = Join-Path $RepoRoot "logs/moderation-$scope.json" |
| 455 | |
| 456 | Write-Verbose "Moderating $($records.Count) model outputs for artifact: $ArtifactId" |
| 457 | try { |
| 458 | & $ModerationScript -Records $records -Scope $scope -Threshold $Threshold -OutFile $outFile -ErrorAction Stop |
| 459 | $moderationExitCode = $LASTEXITCODE |
| 460 | } |
| 461 | catch { |
| 462 | Write-Warning "Content moderation script failed: $_" |
| 463 | return @{ flagged = $true; flaggedCount = $records.Count; outputPath = $outFile } |
| 464 | } |
| 465 | |
| 466 | $flagged = $moderationExitCode -ne 0 |
| 467 | $flaggedCount = 0 |
| 468 | if (Test-Path -LiteralPath $outFile) { |
| 469 | $output = Get-Content -LiteralPath $outFile -Raw | ConvertFrom-Json |
| 470 | $flaggedCount = [int]$output.summary.flaggedCount |
| 471 | } |
| 472 | |
| 473 | return @{ |
| 474 | flagged = $flagged |
| 475 | flaggedCount = $flaggedCount |
| 476 | outputPath = $outFile |
| 477 | } |
| 478 | } |
| 479 | |
| 480 | Export-ModuleMember -Function @( |
| 481 | 'Resolve-VallyRunDir', |
| 482 | 'Read-VallyResultsJsonl', |
| 483 | 'Invoke-VallySpec', |
| 484 | 'Test-SpecInputModeration', |
| 485 | 'Test-SpecOutputModeration' |
| 486 | ) |
| 487 | |