In order to ensure the highest quality model adaptation, be sure that translation memory files consist mostly of full, natural-language sentences. Uploading a large quantity of short segments may degrade model performance and require a model reset.
For on-premise customers, processing of TMs larger than 3,000 parallel segments is only recommended when a GPU is installed on the system. Otherwise, memory updates may be slow.
With a GPU, Lilt processes TMs much faster, although the exact rate varies by language pair, type of GPU, and amount of CPU and RAM available to the system.
The maximum file size for TM uploads is 200 MB. For file sizes larger than 200 MB, customers may ZIP files before upload.