Translation Fine-tuning assistance

A community forum post asks for advice on fine-tuning models for Japanese–English, English–Arabic and Japanese–Arabic translation, and the response recommends freezing a single identical prompt template across all examples, naming the target language in the same position every time, and computing loss only on the translation itself. The reply warns that varying the direction signal teaches the model to guess the target language from the source, which it says is the origin of mixed-language outputs and is worst on Japanese-to-Arabic where the signal is thinnest. It advises scoring a held-out set per direction, counting any wrong-language output via a simple language-ID pass as a failure before checking BLEU or chrF, and selecting the base model by that off-target rate after the same short fine-tune on a small slice for each candidate.

Hi everyone, I am currently working on fine-tuning models for multi-directional translation tasks—specifically focusing on Japanese–English ja-en , English–Arabic en-ar , and Japanese–Arabic ja-ar . To ensure the best possible translation quality and prevent common pitfalls like hallucination or mixed-language outputs , I’m looking to optimize my workflow. Could the community share recommendations on: 1. Dataset Formatting & Templates: What are the most effective prompt structures or chat templates for translation fine-tuning e.g., standard source-target pairs vs. instruction-based formatting ? 2. Instruction Design: Are there specific system prompts or instruction guidelines that yield the best results for low-resource pairs or distant language pairs like Japanese-to-Arabic? 3. Recommended Base Models: Which open-source base models e.g., Llama, Mistral, Qwen, or specific multilingual models have you found to be most robust for these specific language directions? Any advice, recommended templates, or references to successful repositories would be greatly appreciated Thanks in advance For three directions across three scripts, freeze the format before anything else: one identical template for every example, the target language named in the same place every time, and loss computed only on the translation itself. If the direction signal varies, the model learns to guess the target from the source instead — which is where mixed-language outputs start, worst on ja→ar where the signal is thinnest. Check it directly: on a held-out set scored per direction, count any output in the wrong language a simple language-ID pass as a failure before you look at BLEU or chrF, and choose the base model by that off-target rate after the same short fine-tune on a small slice for each candidate.