HERIT: Enabling Global Access to Hanja-to-English Historical Translation by Mitigating Temporal Bias

HERIT addresses the data scarcity and temporal bias inherent in historical archives written in Hanja. It combines high-quality pseudo-labeled data augmentation via retrieval-augmented generation (RAG) with two-stage fine-tuning.

📋 Table of Contents

• Highlights
• Dataset
• Fine-tuned Models
• Model Versions and Access Dates

✨ Highlights

• Mitigates temporal bias: human evaluations stratified by reign confirm that HERIT translates robustly across historical periods not covered by expert-translated data.
• Strong empirical results: HERIT outperforms strong baselines on lexical-overlap metrics and achieves a preference rate of about 66% in human expert evaluation.
• Large-scale pseudo-labeling: 315K RAG-based pseudo-labeled documents for augmentation, and 2.09M final Hanja-to-English translations.
• Full pipeline released: this repository contains the code used to build HERIT — data crawling and preprocessing, pseudo-labeling, fine-tuning, and evaluation.

📊 Dataset

Note

Work in progress; to be updated by early October 2026.

Dataset	# Documents	Released Fields
$D^{train}$	15.7K	(Document ID, Hanja)
$D^{valid}$	1,000	(Document ID, Hanja)
$D^{test}$	1,000	(Document ID, Hanja)
$D^{test}_{NT}$	2,080	(Document ID, Hanja)
Human-eval subset	200	(Document ID, Hanja)
$D^{aug}$	315K	(Document ID, Hanja, Pseudo-English)
Final translations	2.09M	(Document ID, Hanja, Pseudo-English)

🤖 Fine-tuned Models

HERIT model checkpoints (two-stage fine-tuned from Qwen3 8B / 32B). To be released.

Model	Base Model	Link
HERIT-8B	Qwen3-8B	TBA
HERIT-32B	Qwen3-32B	TBA

🗓 Model Versions and Access Dates

Closed-source models (API)

• Gemini-2.5-Flash (gemini-2.5-flash, updated in June 2025)
• Gemini-2.5-Pro (gemini-2.5-pro, updated in June 2025)
• Gemini-Embedding (gemini-embedding-001)
• GPT-5.1 (gpt-5.1-2025-11-13)
• Sonnet-4.5 (claude-sonnet-4-5-20250929)
• Sonnet-4.6 (claude-sonnet-4-6)

Open-weight models

• Qwen3 8B (unsloth/Qwen3-8B, updated May 14, 2025)
• Qwen3 32B (unsloth/Qwen3-32B, updated May 14, 2025)
• Hunyuan-MT-7B (tencent/Hunyuan-MT-7B, updated Sep 18, 2025)
• Qwen3 235B-A22B (Qwen/Qwen3-235B-A22B-Instruct-2507; accessed via Fireworks AI, Aug 2025 – Feb 2026)
• Kimi-K2-Thinking (moonshotai/Kimi-K2-Thinking; accessed via Fireworks AI, Nov 2025 – Feb 2026)
• DeepSeek-V3.1 (deepseek-ai/DeepSeek-V3.1; accessed via Fireworks AI, Aug 2025 – Feb 2026)