#Use Lina and local HY-MT1.5-1.8B to translate localization entries

This guide describes a localization translation practice: deploy a translation-specific small model locally, expose it as an OpenAI-compatible service, and configure it for Lina to translate localization entries in batches.

This approach is suitable for translating many system entries, plugin text, menus, collection titles, and field labels. Compared with online models, local models are not affected by external API RPM, TPM, or concurrency limits, and concurrency can be tuned according to machine and model capability.

#Overview

This guide uses:

• Model: tencent/HY-MT1.5-1.8B-GGUF
• Inference service: llama-server
• Integration: OpenAI-compatible API
• AI Employee: Lina
• Entry point: Localization Management page

#Prerequisites

Before starting, prepare:

• The Localization Management plugin is enabled.
• Target language is enabled.
• Localization entries have been synchronized.
• The local machine or server can run llama-server.
• The NocoBase service can access the HTTP address of llama-server.

#Deploy HY-MT GGUF

#Install llama.cpp

On macOS, you can install it with Homebrew:

brew install llama.cpp

You can also use a prebuilt llama.cpp binary or build it from source. The final requirement is that llama-server is available.

#Start an OpenAI-compatible service

Start the service with the GGUF model from Hugging Face:

llama-server \
  -hf tencent/HY-MT1.5-1.8B-GGUF:Q4_K_M \
  --host 0.0.0.0 \
  --port 8000 \
  -c 2048 \
  -np 4

#Test the Model Service

After llama-server starts, check service health:

curl http://127.0.0.1:8000/health

Then test translation through the OpenAI-compatible API:

curl http://127.0.0.1:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "tencent/HY-MT1.5-1.8B-GGUF:Q4_K_M",
    "messages": [
      {
        "role": "user",
        "content": "Translate the following text into Chinese. Output only the translated result without any additional explanation:\n\nSave"
      }
    ]
  }'

If you start from a local model file, change model to the actual model name returned or configured by the service.

#Configure an LLM Service in NocoBase

Go to System Settings -> AI Employees -> LLM service and add an LLM service.

Example configuration:

After configuration, use Test flight to verify the model.

#Configure Lina's Dedicated Model

Go to System Settings -> AI Employees -> AI employees, open Lina, and switch to Model settings.

1. Enable Enable dedicated model configuration.
2. Select the HY-MT local model in Models.
3. Save the configuration.

After this, Lina uses this model for localization translation tasks, preventing users or tasks from switching to general chat models.

For details, see Configure AI Employee Models.

#Configure Translation Concurrency

Localization translation task concurrency is controlled by AI_LOCALIZATION_CONCURRENCY:

AI_LOCALIZATION_CONCURRENCY=10

Rules:

• Default: 10
• Minimum: 1
• Maximum: 20
• Values outside the range use the default

The best concurrency depends on CPU, GPU, memory, model quantization, and llama-server -np. If the default concurrency causes issues:

1. Start with AI_LOCALIZATION_CONCURRENCY=1 and verify single-entry translation.
2. Set both llama-server -np and AI_LOCALIZATION_CONCURRENCY to 2 or 4.
3. Observe response time, CPU/GPU usage, and task progress.
4. Increase concurrency gradually only if stable.

#Execute Localization Translation

Go to System Management -> Localization Management.

1. Switch to the target language.
2. Click Synchronize to ensure entries are synchronized.
3. Click Lina's avatar.
4. Choose a task scope:
   • Incremental translation: translate entries without translations.
   • Selected translation: translate selected entries in the table.
   • Full translation: translate all entries in the current language.
5. Check entry count, provider, and model in the confirmation dialog.
6. Confirm to create the async task.
7. Wait for completion, review translations, and publish.

Start with Selected translation for a few entries to verify output style and speed before running incremental or full translation.

#How Lina Builds Translation Requests

Lina builds requests from entries and reference translations. For short entries, existing references are used to improve consistency:

• Built-in entries prefer Chinese translations as references.
• Non-built-in entries prefer the system default language as references.
• If an English reference exists, English is used as source text.
• Translation results are written to the target language but are not published automatically.

Prompt semantics are similar to:

Refer to the following translation:
{source_term} is translated as {target_term}

Translate the following text into {target_language}. Output only the translated result without any additional explanation:

{source_text}

#Troubleshooting

#No progress after creating a task

Check whether llama-server received requests. View service logs or call /v1/chat/completions with curl.

If the model receives requests but does not return, reduce:

• AI_LOCALIZATION_CONCURRENCY
• llama-server -np
• llama-server -c

#The model returns explanations instead of translations

Local translation models are usually more stable than general chat models. If explanations still appear, test the same prompt with curl first to verify the model's output style.

You can also translate shorter entries first or reduce sampling parameters such as temperature.

#NocoBase cannot connect to the model service

Check:

• Whether Base URL includes /v1.
• Whether the NocoBase runtime environment can access the address.
• Whether firewall or container networking blocks the port.
• Whether llama-server is still running.

#Review Before Publishing

After AI translation finishes, review before publishing:

• Filter by module and check short entries such as menus, buttons, field names, and statuses.
• Check variables, placeholders, HTML tags, and formatting symbols.
• Check key business terminology consistency.
• If built-in entry translations are overwritten, resynchronize in Localization Management and select Reset system built-in entry translations to restore defaults. To contribute default translations for the system and official plugins, see Translation Contribution.
• Publish in a test environment first, then sync to production.

#References

• tencent/HY-MT1.5-1.8B-GGUF
• llama-server documentation
• Lina: Localization Engineer