Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HY-MT1.5 Technical Report

Mao Zheng, Zheng Li, Tao Chen, Mingyang Song, Di Wang

TL;DR

HY-MT1.5 tackles the quality-efficiency trade-off and customization in machine translation by presenting a holistic training framework that combines MT-oriented pretraining, supervised fine-tuning, on-policy distillation, and reinforcement learning. The 1.8B and 7B variants demonstrate strong parameter efficiency, outperforming many open-source baselines and commercial APIs, while approaching or surpassing ultra-large proprietary models on multiple benchmarks; the approach uses a rubrics-based reward and strong-to-weak distillation to transfer performance from the larger 7B model. Practical features such as terminology intervention, context-aware translation, and format preservation extend real-world applicability, and quantization experiments show deployment viability with FP8 and PTQ approaches. The results suggest a scalable path to high-quality, efficient, and customizable MT suitable for edge devices and high-throughput translation tasks.

Abstract

In this report, we introduce our latest translation models, HY-MT1.5-1.8B and HY-MT1.5-7B, a new family of machine translation models developed through a holistic training framework tailored for high-performance translation. Our methodology orchestrates a multi-stage pipeline that integrates general and MT-oriented pre-training, supervised fine-tuning, on-policy distillation, and reinforcement learning. HY-MT1.5-1.8B, the 1.8B-parameter model demonstrates remarkable parameter efficiency, comprehensively outperforming significantly larger open-source baselines (e.g., Tower-Plus-72B, Qwen3-32B) and mainstream commercial APIs (e.g., Microsoft Translator, Doubao Translator) in standard Chinese-foreign and English-foreign tasks. It achieves approximately 90% of the performance of ultra-large proprietary models such as Gemini-3.0-Pro, while marginally trailing Gemini-3.0-Pro on WMT25 and Mandarin-minority language benchmarks, it maintains a substantial lead over other competing models. Furthermore, HY-MT1.5-7B establishes a new state-of-the-art for its size class, achieving 95% of Gemini-3.0-Pro's performance on Flores-200 and surpassing it on the challenging WMT25 and Mandarin-minority language test sets. Beyond standard translation, the HY-MT1.5 series supports advanced constraints, including terminology intervention, context-aware translation, and format preservation. Extensive empirical evaluations confirm that both models offer highly competitive, robust solutions for general and specialized translation tasks within their respective parameter scales.

HY-MT1.5 Technical Report

TL;DR

HY-MT1.5 tackles the quality-efficiency trade-off and customization in machine translation by presenting a holistic training framework that combines MT-oriented pretraining, supervised fine-tuning, on-policy distillation, and reinforcement learning. The 1.8B and 7B variants demonstrate strong parameter efficiency, outperforming many open-source baselines and commercial APIs, while approaching or surpassing ultra-large proprietary models on multiple benchmarks; the approach uses a rubrics-based reward and strong-to-weak distillation to transfer performance from the larger 7B model. Practical features such as terminology intervention, context-aware translation, and format preservation extend real-world applicability, and quantization experiments show deployment viability with FP8 and PTQ approaches. The results suggest a scalable path to high-quality, efficient, and customizable MT suitable for edge devices and high-throughput translation tasks.

Abstract

In this report, we introduce our latest translation models, HY-MT1.5-1.8B and HY-MT1.5-7B, a new family of machine translation models developed through a holistic training framework tailored for high-performance translation. Our methodology orchestrates a multi-stage pipeline that integrates general and MT-oriented pre-training, supervised fine-tuning, on-policy distillation, and reinforcement learning. HY-MT1.5-1.8B, the 1.8B-parameter model demonstrates remarkable parameter efficiency, comprehensively outperforming significantly larger open-source baselines (e.g., Tower-Plus-72B, Qwen3-32B) and mainstream commercial APIs (e.g., Microsoft Translator, Doubao Translator) in standard Chinese-foreign and English-foreign tasks. It achieves approximately 90% of the performance of ultra-large proprietary models such as Gemini-3.0-Pro, while marginally trailing Gemini-3.0-Pro on WMT25 and Mandarin-minority language benchmarks, it maintains a substantial lead over other competing models. Furthermore, HY-MT1.5-7B establishes a new state-of-the-art for its size class, achieving 95% of Gemini-3.0-Pro's performance on Flores-200 and surpassing it on the challenging WMT25 and Mandarin-minority language test sets. Beyond standard translation, the HY-MT1.5 series supports advanced constraints, including terminology intervention, context-aware translation, and format preservation. Extensive empirical evaluations confirm that both models offer highly competitive, robust solutions for general and specialized translation tasks within their respective parameter scales.
Paper Structure (16 sections, 4 figures, 4 tables)

This paper contains 16 sections, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. MT-oriented Pretraining and Supervised Fine-Tuning
  4. Reinforcement Learning
  5. Strong-to-Weak On-Policy Distillation
  6. Quantization
  7. Experiments
  8. Automatic Metrics
  9. Human Evaluation
  10. Efficiency of HY-MT1.5 Models
  11. Practical Distinctive Features
  12. Quantization Experiment
  13. Conclusion
  14. Contributions
  15. Core Contributors
  16. ...and 1 more sections

Figures (4)

  • Figure 1: Benchmark performance of HY-MT1.5 models and state-of-the-art baselines.
  • Figure 2: Comparison of model size versus translation quality across Flores-200 and WMT25 datasets for open-source models.
  • Figure 3: Training pipline of HY-MT1.5-1.8B and HY-MT1.5-7B.
  • Figure 4: Average response time versus translation quality for different translation models.