TranslateEval

LLM-powered translation quality assessment for game localization and TQA workflows.

Evaluate source-target text pairs with scored grading, batch CSV processing, and exportable reports — using your own LLM provider (local or cloud).

---

Features

• 📊 0–100 scoring across Accuracy, Fluency, Terminology, Style
• 📋 Batch CSV mode — upload a source,target file, get a full report
• 🏷️ 4-tier classification — Publish-Ready / Flawed / Unusable / Won't Process
• ☁️ OpenAI-compatible APIs — OpenAI, DashScope, Groq, DeepSeek, SiliconFlow
• 🔷 Anthropic Claude — native SDK support
• ⚡ Local Ollama — run fully offline with qwen2.5:14b or any local model
• 📥 CSV export — per-segment breakdown with issues and suggestions

Scoring Tiers

Score	Tier	Meaning
91–100	✅ Publish-Ready	Ready for release without edits
70–90	⚠️ Flawed	Notable errors requiring correction
50–69	❌ Unusable	Critical errors; retranslation recommended
0–49	🚫 Won't Process	Untranslatable or placeholder content

Quick Start

git clone https://github.com/Grayson0130/translateeval.git
cd translateeval
npm install
npm run dev
# → http://localhost:3000

Supported Providers

Mode	Provider	Example model
⚡ Local	Ollama	qwen2.5:14b
☁️ OpenAI fmt	OpenAI	gpt-4o-mini
☁️ OpenAI fmt	DashScope	qwen-plus
☁️ OpenAI fmt	Groq	llama-3.1-8b-instant
☁️ OpenAI fmt	DeepSeek	deepseek-chat
🔷 Anthropic	Claude	claude-3-5-haiku-20241022

CSV Format

source,target
"Original text","Translated text"

Tech Stack

Next.js 14 · TypeScript · openai SDK · @anthropic-ai/sdk · papaparse

License

MIT © Jinhui Zhao