Practical overview
When it fits
Use Drt when you need research and model resources for experimenting with reasoning-oriented machine translation, rather than a ready-made hosted translation service.
Who it is for
- Researchers studying reasoning-based machine translation.
- Developers who want to inspect or run the linked translation checkpoints.
- Readers exploring multilingual translation, reinforcement learning or reasoning-oriented translation data.
Documented uses
- Find DRT, DeepTrans, ExTrans and mExTrans model resources.
- Explore research on long chain-of-thought translation, reinforcement learning and exemplar-enhanced multilingual translation.
- Follow documented examples to generate translations with Transformers or serve a checkpoint through vLLM.
- Access released DRT data, including synthesized thought content and translations.
Requirements & limits
The repository is public and links to model checkpoints on Hugging Face. Its README documents self-managed inference workflows.
- DRT-7B, DRT-8B and DRT-14B are linked to Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct and Qwen2.5-14B-Instruct backbones, respectively.
- The README’s DRT Transformers example translates English text into Chinese.
- The repository links related DeepTrans and ExTrans model resources, with separate inference examples for some models.
- The README describes a three-agent synthesis framework and reports 22,264 synthesized DRT samples.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Generate a translation with DRT-7B and Transformers
Load DRT-7B, provide a translation prompt and decode a generated response using the README’s Transformers example.
Before you start
- Python with the Transformers and PyTorch libraries used by the example.
- Access to the Krystalan/DRT-7B model checkpoint and tokenizer.
Steps
Load AutoModelForCausalLM and AutoTokenizer from Transformers using the Krystalan/DRT-7B model name.
Write a translation prompt and format it as chat messages with the tokenizer’s chat template.
Tokenize the formatted prompt, move the inputs to the model device and call model.generate.
Remove the input tokens from the generated sequence, decode the response and print it.
Expected outcome
A decoded model response to the translation prompt; the README example uses English-to-Chinese text.
Guide limitations
- The README does not specify hardware requirements or guarantee compatibility with a particular environment.
- The example demonstrates a prompt and generated response, not a hosted translation service.
Serve a translation checkpoint with vLLM
Follow the documented pattern for serving a selected checkpoint with vLLM and submitting a translation request through an OpenAI client.
Before you start
- A model checkpoint and model name to provide to the documented vLLM server command.
- vLLM and the OpenAI Python client, as used in the README examples.
Steps
Start the vLLM OpenAI-compatible API server with the selected checkpoint and served model name.
Configure the OpenAI client with the documented local API base and key settings.
Submit a chat completion request containing a translation prompt and the served model name.
Expected outcome
A chat completion response from the locally served translation model.
Guide limitations
- The example uses a local server address and requires the server to be running.
- The repository does not specify hardware requirements or guarantee compatibility.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- krystalan/DRT · Official homepage title · retained in its source language