EN→JA

Terminology-aware English-to-Japanese translation. Comparing prompt-only, fine-tuned, RAG, and agentic approaches.

Try the Translator View Research
S0
Baseline
Prompt-only Qwen 2.5-0.5B. No adaptation.
S1
Fine-Tuned
QLoRA adapters on Qwen. Task-specific training.
S2
RAG + Reranker
Multi-strata retrieval with cross-encoder reranking.
S3
Agentic RAG
Claude Sonnet 4.6 with ReAct, tool use, and self-audit.
Datasets & Knowledge Base
📚
Training Data
24K EN-JA pairs
Tatoeba, JParaCrawl
📑
Glossary
320+ terms
Tech/UI terminology
📋
KB Chunks
47K+ embedded
Parallel, grammar, style, errors
🔍
Test Set
250-4,254
Varies by variant
Data Pipeline
Collect → Normalize → Filter & Dedup → Split → Chunk → Embed → S3 Vectors
Tech Stack
Base Model
Qwen 2.5-0.5B
Agentic Model
Claude Sonnet 4.6
Vector Store
AWS S3 Vectors
Embeddings
multilingual-e5-small