0.8B Translation Model Beats 1.4B Frontier via MO-GRPO and Glossary Compliance
Nihongi-studier12387 · reddit · 2026-08-14
A developer shared how they fine-tuned a 0.8B parameter JA↔EN translation model (CyberAgent CAT-Translate) to outperform the stock 1.4B version.
- Goal: Enable the model to obey a per-word glossary at inference, optimized for mobile phones.
- Training: Stage 1 used LoRA SFT to teach the model to prioritize supplied glosses. Stage 2 applied MO-GRPO (Ichihara et al. 2025), which normalizes multi-objective rewards (translation quality and glossary compliance) independently to prevent reward collapse seen in vanilla GRPO.
- Performance: In EN→JA business dialogues, the trained 0.8B model beats the stock 1.4B across COMET-QE, MetricX-24, chrF, and BLEU. It achieves 91% glossary adherence and runs at 103 ms/sentence in int8 on Apple silicon.
More from Research
- Cooperative AI Seminar: Solving AI Game Theory Dilemmas with Safe Pareto Improvements — xuanalogue · 2026-08-14
- AI Brain Diagnostic Startup Hemispheric Raises $52M — rjhaier · 2026-08-14
- Roundup of RVQ Codec Research and Workarounds for Audio Models — andrew_n_carr · 2026-08-14
- Search Agent Evals Inflated: Models Cheat by Querying Benchmark Answers on GitHub — scaling01 · 2026-08-14
- Why LLM RL Works: Exploring Low-Bias Methods Despite Information-Theoretic Inefficiency — agarwl_ · 2026-08-14
- From Output Auditing to Internal Representations: The Value of Anthropic's Interpretability Research — krishnan · 2026-08-14