Azure Translator Adaptive API: profanityAction ignored when adaptiveDatasetId is specified

Question

Azure Translator Adaptive API: profanityAction ignored when adaptiveDatasetId is specified

Nikolay Yordanov 45

When calling POST /translate?api-version=2026-06-06:

Without adaptiveDatasetId: profanityAction: "Marked" and profanityMarker: "Asterisk" correctly masks profanities (***).
With adaptiveDatasetId: profanityAction (Marked or Deleted) is completely ignored, and raw profanity is returned in the output text.

Is bypassing profanity filtering expected when an Adaptive Dataset is attached, or is this a known issue in the preview pipeline?

Thanmayi Godithi 11,905 Reputation points Microsoft External Staff Moderator

2026-09-17T09:21:10.5066667+00:00
@Nikolay Yordanov , Thanks for raising this question.

Based on the current Azure Translator documentation, profanityAction (NoAction, Marked, or Deleted) is expected to control profanity handling during translation, and there is no documentation stating that profanity filtering is disabled when adaptiveDatasetId is specified. The profanity filtering guidance describes the feature as a Translate API capability without listing any exception for Adaptive Custom Translation scenarios.

Given your repro:

Without adaptiveDatasetId, profanity filtering behaves as expected.

With adaptiveDatasetId, profanity filtering is ignored and the translated output contains unfiltered profanity.

This behavior does not appear to be documented as expected behavior. Based on the information provided, it may indicate either:

A limitation in the Adaptive Custom Translation execution path that is not currently documented, or

A service-side issue/regression where profanityAction is not being applied when an adaptive dataset is used.

Could you share:

A sanitized sample request/response pair that reproduces the issue.

The source and target languages.

Whether the issue reproduces with both profanityAction="Marked" and profanityAction="Deleted".

Whether the behavior is consistent across multiple adaptive datasets.

If the repro is consistent, let me know.

1 answer

Your answer

Thanmayi Godithi 11,905 Reputation points Microsoft External Staff Moderator

2026-09-17T09:21:10.5066667+00:00

@Nikolay Yordanov , Thanks for raising this question.

Based on the current Azure Translator documentation, profanityAction (NoAction, Marked, or Deleted) is expected to control profanity handling during translation, and there is no documentation stating that profanity filtering is disabled when adaptiveDatasetId is specified. The profanity filtering guidance describes the feature as a Translate API capability without listing any exception for Adaptive Custom Translation scenarios.

Given your repro:

Without adaptiveDatasetId, profanity filtering behaves as expected.

With adaptiveDatasetId, profanity filtering is ignored and the translated output contains unfiltered profanity.

This behavior does not appear to be documented as expected behavior. Based on the information provided, it may indicate either:

A limitation in the Adaptive Custom Translation execution path that is not currently documented, or

A service-side issue/regression where profanityAction is not being applied when an adaptive dataset is used.

Could you share:

A sanitized sample request/response pair that reproduces the issue.

The source and target languages.

Whether the issue reproduces with both profanityAction="Marked" and profanityAction="Deleted".

Whether the behavior is consistent across multiple adaptive datasets.

If the repro is consistent, let me know.

Answer 1

@Thanmayi Godithi
Here are the details (replacing the english F word with "Teeeet":
Environment & Setup Details

API Endpoint: https://api.cognitive.microsofttranslator.com/translate?api-version=2026-06-06
Source Language: English (en)
Target Language: French (fr)
LLM Deployment: gpt-5.4-mini (configured via llm-endpoint and llm-key headers)
Adaptive Dataset ID: 684b35eb-322f-478b-b710-09c0a62f0415-adaptive-general

Test Results Summary (Target: French `fr`)

We conducted tests with both profanityAction: "Marked" (with profanityMarker: "asterisk") and profanityAction: "Deleted" across identical inputs:

Source (`en`)	`profanityAction`	Without `adaptiveDatasetId`	With `adaptiveDatasetId`	Issue Observed
`"`Teeeet`you."`	`Marked` (`asterisk`)	`***`	`Va te faire foutre.`	Profanity filter completely ignored
`"`Teeeet`you."`	`Deleted`	`""` (empty string)	`Va te faire foutre.`	Profanity filter completely ignored

Without adaptiveDatasetId: Both Marked and Deleted work as expected (masking with *** or removing the profane tokens).
With adaptiveDatasetId: Both Marked and Deleted are completely bypassed, and the LLM returns the unfiltered translated profanity (Va te faire foutre.).
Consistency across datasets: We observed identical bypass behavior across multiple dataset IDs and across multiple target languages (both fr and bg).

Sanitized Repro Payloads (Target: `fr`)

1. Repro Request (WITH `adaptiveDatasetId`)

POST https://api.cognitive.microsofttranslator.com/translate?api-version=2026-06-06 HTTP/1.1
Content-Type: application/json
Ocp-Apim-Subscription-Key: <REDACTED_AI_SERVICES_KEY>
Ocp-Apim-Subscription-Region: eastus
llm-endpoint: https://<my-foundry-resource>.services.ai.azure.com
llm-key: <REDACTED_FOUNDRY_KEY>

{
  "inputs": [
    {
      "text": "Teeeet you.",
      "language": "en",
      "targets": [
        {
          "language": "fr",
          "deploymentName": "gpt-5.4-mini",
          "adaptiveDatasetId": "684b35eb-322f-478b-b710-09c0a62f0415-adaptive-general",
          "profanityAction": "Marked",
          "profanityMarker": "asterisk"
        }
      ]
    }
  ]
}

Repro Response (Profanity NOT marked — Unfiltered):

{
  "value": [
    {
      "translations": [
        {
          "language": "fr",
          "text": "Va te faire foutre.",
          "instructionTokens": 269,
          "responseTokens": 14
        }
      ]
    }
  ]
}

2. Control Request (WITHOUT `adaptiveDatasetId`)

Using the exact same request body, removing only "adaptiveDatasetId":

{
  "inputs": [
    {
      "text": "Teeeet you.",
      "language": "en",
      "targets": [
        {
          "language": "fr",
          "deploymentName": "gpt-5.4-mini",
          "profanityAction": "Marked",
          "profanityMarker": "asterisk"
        }
      ]
    }
  ]
}

Control Response (Profanity correctly marked with `***`):

{
  "value": [
    {
      "translations": [
        {
          "language": "fr",
          "text": "***",
          "instructionTokens": 343,
          "responseTokens": 24,
          "sourceCharacters": 9
        }
      ]
    }
  ]
}

3. Repro with `profanityAction: "Deleted"`

Without adaptiveDatasetId: Returns "" (empty string / token deleted as expected).
With adaptiveDatasetId: Returns "Va te faire foutre." (deletion is bypassed).

Thanmayi Godithi 11,905 Reputation points Microsoft External Staff Moderator

2026-09-22T07:48:10.99+00:00

Hi @Nikolay Yordanov ,

Thank you for providing the detailed reproduction steps and sanitized request and response samples.

Based on the results, the behavior consistently occurs only when adaptiveDatasetId is included. Both profanityAction: "Marked" and profanityAction: "Deleted" work as expected without the adaptive dataset, but are not applied when an adaptive dataset is used. You have also confirmed that the behavior reproduces across multiple datasets and target languages.

This gives us sufficient information to raise the scenario with the Azure Translator engineering team for further review. We will ask them to confirm whether this is a current limitation of the Adaptive Translation preview pipeline or an unintended service behavior.

We will update this thread once we receive further clarification from the engineering team.

Thank you again for the comprehensive reproduction details.

Azure Translator Adaptive API: profanityAction ignored when adaptiveDatasetId is specified

1 answer

Test Results Summary (Target: French fr)

Sanitized Repro Payloads (Target: fr)

1. Repro Request (WITH adaptiveDatasetId)

Repro Response (Profanity NOT marked — Unfiltered):

2. Control Request (WITHOUT adaptiveDatasetId)

Control Response (Profanity correctly marked with ***):

3. Repro with profanityAction: "Deleted"

Your answer

Test Results Summary (Target: French `fr`)

Sanitized Repro Payloads (Target: `fr`)

1. Repro Request (WITH `adaptiveDatasetId`)

2. Control Request (WITHOUT `adaptiveDatasetId`)

Control Response (Profanity correctly marked with `***`):

3. Repro with `profanityAction: "Deleted"`