|
--- |
|
datasets: |
|
- RUCAIBox/Erya-dataset |
|
language: |
|
- en |
|
base_model: |
|
- google/gemma-2-2b-it |
|
tags: |
|
- ancient-chinese |
|
- chinese |
|
- literature |
|
- unsloth |
|
- trl |
|
- sft |
|
--- |
|
Ancient Chinese Translator + Phonology Model (SimaQian) |
|
|
|
Name Origin: |
|
|
|
The origin of the model name comes from famous ancient chinese historian Qian Sima (司馬遷), known for his Records of the Grand Historian, a general history of China covering more than two thousand years. |
|
|
|
This model combines two key functionalities for Ancient Chinese texts: |
|
|
|
1. Translation: Converts Ancient Chinese passages into modern Chinese. |
|
|
|
2. Phonological Reconstruction: Provides historical pronunciations for characters or entire sentences across multiple eras (e.g., Middle Tang, Song, Yuan, Ming/Qing). |
|
|
|
|
|
Model Description |
|
|
|
• Architecture: Fine-tuned on top of Google’s Gemma 2 model using LoRA. |
|
|
|
• Input Format: Special tokens <start_of_turn> / <end_of_turn> define user vs. model turns. |
|
|
|
• Output: Era identification (optional), phonetic renderings, and modern Chinese translations. |
|
|
|
Training Data |
|
• Translation: Erya dataset from RUCAIBox/Erya-dataset. |
|
• Phonology: Ancient-Chinese-Phonology (ACP) for multi-era reconstructions. |
|
• Fine-Tuning: LoRA-based parameter-efficient approach on Gemma 2 Instruct. |
|
|
|
|
|
Usage |
|
|
|
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM |
|
|
|
tokenizer = AutoTokenizer.from_pretrained("lordChipotle/SimaQian") |
|
|
|
model = AutoModelForCausalLM.from_pretrained("lordChipotle/SimaQian") |
|
|
|
|
|
prompt = """ |
|
<start_of_turn>user |
|
Given the ancient text: 「子曰：學而時習之，不亦說乎？」 |
|
1) Identify the era |
|
2) Provide the phonetic reading |
|
3) Translate into modern Chinese |
|
<end_of_turn> |
|
<start_of_turn>model |
|
""" |
|
|
|
inputs = tokenizer(prompt, return_tensors="pt") |
|
|
|
outputs = model.generate(**inputs, max_length=256) |
|
|
|
print(tokenizer.decode(outputs[0])) |
|
|
|
|
|
Limitations and Biases |
|
|
|
• Era Estimation: Model may not always correctly guess the historical era. |
|
|
|
• Pronunciations: Reconstructions are approximate and can vary by scholarly consensus. |
|
|
|
• Contextual Accuracy: For highly contextual Ancient Chinese passages, translations may need further review by domain experts. |
|
|
