Inquiries on Azure Adaptive Custom Translator (AdaptCT)

Question

I have a few questions regarding AdaptCT that I cannot seem to find from the documents:

1. Is it possible to do incremental update of dataset? Currently I can only see the import API for dataset but what if I just want to append something to the existing document?
2. Is there a size limit in the files uploaded as documents to be used in the dataset? Is there a character or line limit per file?
3. Are there cost implications when creating and updating datasets? Does the size of the datasets has cost impact?
4. Any recommendations on how to build the dataset (given that the context matters to produce good translation)? How do we build and maintain the dataset to keep it updated and relevant?

Answer 1

Hello @Maricar Hernandez

Thank you for reaching out with your questions regarding Azure Adaptive Custom Translator (AdaptCT).

I understand you're looking for guidance on dataset management, operational limits, costs, and best practices for maintaining high-quality adaptive translation datasets.

Please find responses below:

1. Is incremental update of an existing dataset supported?

Adaptive Custom Translator is designed around a developer-managed adaptive dataset lifecycle through the AdaptCT API (v1.0 Preview). The Adaptive Custom Translation Playground also provides a no-code experience for managing datasets.

Based on the currently available documentation:

The service supports dataset lifecycle management, including importing and managing datasets.

However, the documentation does not explicitly state whether an existing dataset can be updated by appending new content, or whether updates require replacing or rebuilding the dataset.

The available Import API is intended for dataset management, but append semantics are not documented.

At this time, we cannot confirm that incremental append operations are supported. If your scenario requires adding new translation pairs over time without recreating the dataset, this would require confirmation from the Product Group.

2. Is there a size limit for uploaded documents?

Based on the available documentation:

• Maximum upload file size: 100 MB
• Files larger than this are automatically split into smaller parts during processing.

For the uploaded content:

• Individual source or target segments must not exceed 250 characters.
• Segments exceeding this limit are rejected.
• If all segments in a document are invalid, the document upload fails.
• If an upload fails because of invalid segments, the Get Import Job Status API can be used to identify the offending segments.

For AdaptCT datasets specifically, Microsoft recommends:

• Uploading approximately 5–10,000 pre-aligned source/target segment pairs
• Keeping each segment 250 characters or fewer

The current documentation does not specify any limit on:

• Number of lines per file
• Total characters per file (other than the 100 MB upload limit)

3. Are there costs associated with creating or updating datasets?

Based on the documentation currently available, Microsoft has not published pricing details specific to:

• Creating AdaptCT datasets
• Updating existing datasets
• Whether dataset size directly affects pricing

Therefore, we cannot confirm whether dataset creation or updates incur additional charges beyond the standard service pricing.

Any Azure Storage used to store your source documents would continue to follow the standard Azure Storage pricing model.

For pricing specific to AdaptCT Preview, we recommend referring to the official Azure AI Translator pricing documentation as additional guidance becomes available.

4. Best practices for building and maintaining datasets

Adaptive Custom Translator differs from traditional Custom Translator because it does not rely on offline model training.

Instead, AdaptCT:

• Creates an adaptive dataset index
• Uses few-shot retrieval during inference
• Retrieves translation examples that are most similar to each incoming request to improve terminology, style, and contextual consistency

To achieve the best translation quality,

• Use high-quality, domain-specific parallel text
• Ensure source and target sentences are correctly aligned
• Keep each translation segment within the 250-character limit
• Build datasets using approximately 5,000–10,000 aligned segment pairs where possible
• Preserve contextual translations rather than isolated phrases whenever practical
• Remove duplicate, outdated, or low-quality translation pairs
• Refresh datasets periodically with newly validated translations to keep terminology current
• Consider maintaining separate datasets for different business domains (for example, legal, medical, financial, or technical content)

One advantage of AdaptCT is that dataset updates are processed within minutes, allowing terminology and style to be refreshed much faster than traditional custom model training.

Since inference retrieves similar examples from your adaptive dataset, the overall quality, coverage, and relevance of your translation pairs have a significant impact on translation accuracy and consistency.

https://azure.microsoft.com/en-in/pricing/details/translator/?utm_source=chatgpt.com

https://devblogs.microsoft.com/foundry/translation-customization-a-developers-guide-to-adaptive-custom-translation/

https://learn.microsoft.com/en-us/azure/ai-services/translator/foundry/adaptive-custom-translation

I Hope this helps. Do let me know if you have any further queries.

---

If this answers your query, please do click Accept Answer and Yes for was this answer helpful.

Thank you!

Answer 2

Hello Maricar Hernandez,

Welcome to the Microsoft Q&A and thank you for posting your questions here.

I understand that your enquiries about Azure Adaptive Custom Translator (AdaptCT):

Regarding your questions:

Is it possible to do incremental update of dataset? Currently I can only see the import API for dataset but what if I just want to append something to the existing document?

No, not in place through the currently documented API. - https://learn.microsoft.com/en-us/azure/ai-services/translator/foundry/adaptive-custom-translation

The AdaptCT API does not provide an operation to:

• Append rows to an existing uploaded document
• Add documents to an existing dataset index
• Update an existing dataset index
• Patch individual source-target pairs
• Delete individual source-target pairs from an index

The supported update method is to import the new or corrected bilingual content and create a new adaptive dataset index.

Is there a size limit in the files uploaded as documents to be used in the dataset? Is there a character or line limit per file?

Yes, there are limits.

Use 5 to 10,000 aligned source-target pairs. Each source and target segment must be no more than 250 characters. A separate MB or per-file line-count limit is not currently published. - https://learn.microsoft.com/en-us/azure/ai-services/translator/foundry/adaptive-custom-translation

Are there cost implications when creating and updating datasets? Does the size of the datasets has cost impact?

No separate public AdaptCT import, indexing, update, storage, or dataset-size price is currently listed. Do not apply traditional Custom Translator training and hosting prices without confirmation. Validate the actual Azure meter and obtain a Microsoft quote for production budgeting. - https://learn.microsoft.com/en-us/azure/ai-services/translator/foundry/adaptive-custom-translation, and https://azure.microsoft.com/en-us/pricing/details/translator/ gives you more insight.

Any recommendations on how to build the dataset (given that the context matters to produce good translation)? How do we build and maintain the dataset to keep it updated and relevant?

Use human-approved, domain-specific, non-conflicting bilingual pairs. Maintain the corpus externally, create versioned indexes, evaluate every new version, switch the application to the new dataset ID, and retain the previous index temporarily for rollback. - https://learn.microsoft.com/en-us/azure/ai-services/translator/foundry/adaptive-custom-translation

I hope this is helpful. Please! Do not hesitate to let me know if you have any other questions, steps or clarifications.

---

Please do not close the thread by upvoting and accepting the answer if any part of it is helpful.

Answer 3

AdaptCT does not expose an append or update operation for an existing dataset index. Import the additional TMX or TSV as a new document, then create a new index that includes the existing and new document IDs. OverwriteIfExists replaces a named document; it is not a row-level append.

Current limits are 5–10,000 aligned source-target segment pairs, with each source and target segment no longer than 250 characters. Microsoft does not publish a separate per-file byte or line limit, so use the segment constraints and check import-job status for rejected segments.

Keep pairs clean, correctly aligned, domain-relevant, and representative of terminology and style. Remove duplicates and conflicts; version source files and indexes, then validate a fixed test set after each rebuild. Documentation does not state a separate dataset-creation charge. Runtime LLM translation is token-billed; confirm current meters in Cost Management or with Azure billing support before production use.