Shiyu-Qwen2.5-14B

Shiyu-Qwen2.5-14B

Model Overview

Shiyu-Qwen2.5-14B is a fine-tuned large language model specialized in Classical Chinese Poetry (CCP) translation and metaphor-aware implicit semantic understanding. It is built upon the Qwen2.5-14B-Instruct base model and enhanced through the MALLM-ISET (Metaphor-Aware LLM Framework for Implicit Semantic-Enhanced Translation) framework.

Unlike conventional translation models that rely on surface-level literal matching, Shiyu-Qwen2.5-14B is designed to perceive and translate conceptual metaphors in classical Chinese poetry—where meaning is rarely encoded in surface lexical forms but embedded in metaphorical mappings that shift dramatically across contexts.

Key Capabilities

• Metaphor-aware translation: Distinguishes between literal and non-literal meanings of polysemous metaphor words (e.g., 青丝 → "green velvet" vs. "music wire" vs. "black hair") based on contextual evidence
• Context-sensitive knowledge retrieval: Integrates a structured Metaphor Knowledge Graph (PMKG) with 77,986 nodes and 290,554 relationships, built from authoritative lexicons
• Reinforcement learning optimization: Leverages Group Relative Policy Optimization (GRPO) with metaphor-word-level reward functions to explicitly distinguish metaphorical, literal, and omitted translation states

---

Model Architecture

Shiyu-Qwen2.5-14B extends the Qwen2.5-14B-Instruct architecture with three core enhancement modules:

Training Pipeline

1. Supervised Fine-Tuning (SFT): LoRA-based parameter-efficient fine-tuning on 66,058 ancient-modern Chinese parallel sentence pairs spanning 15 dynasties
2. Metaphor Knowledge Injection: MA-KGCN-based retrieval with edge weights (α = 0.8 for co-occurrence, β = 0.2 for poem containment)
3. Reinforcement Learning: GRPO optimization with 4 generations per prompt, learning rate of 1e-5, and KL penalty coefficient of 0.04

---

Training Data

Classical Chinese Poetry Corpus

• 66,058 ancient-modern Chinese parallel sentence pairs
• 11,277 unique poems with vernacular translations
• Coverage across 15 dynasties (from Qin to Qing)
• Sourced from chinese-poetry-and-prose, Baidu Baike, and Guwen Island

Metaphor Knowledge Graph (PMKG)

• 13,114 metaphor entries from Thesaurus Dictionary of Synonymous Categories in Classical Chinese Poetry (毛汉华 & 汤春发, 2013)
• 77,986 nodes and 290,554 relationships
• Ontology structure: Metaphor Word → Metaphor Meaning → Metaphor Category, with Co-occur and AppearsIn relations

---

Performance

Sentence-Level Metrics

Metaphor-Specific Evaluation

Linguistic & Literary Evaluation (LLM-as-Judge)

Evaluation framework based on Halliday & Matthiessen (2013), Vinay & Darbelnet (1995), and Newmark (1981).

---

Usage

Installation

pip install transformers torch accelerate

Basic Inference

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "ricardozhy/Shiyu-Qwen2.5-14B"

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

# Example: Translating classical Chinese poetry
prompt = """你是一位中国古典文学高级翻译专家与文献学学者。
请将以下古诗翻译为现代白话文，注意识别并准确翻译其中的隐喻词汇。

古诗：飗飗青丝上，静听松风寒。

翻译："""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.9,
    top_k=50,
    do_sample=True
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Expected Output

For the input poem 飗飗青丝上，静听松风寒。, the model should correctly identify 青丝 as "music wire" (non-literal) rather than "green velvet" (literal), producing a translation similar to:

琴弦奏清凉的曲调，细细倾听就像松涛声。
(The music wire plays a cool, melodious tune, and it sounds just like the rustling of the pine trees.)

---

License

This model is released under the same license as the Qwen2.5 base model. Please refer to the Qwen license for details.

---

Related Resources

• Shiyu Online Translation System — Interactive platform for classical poetry metaphor analysis and translation
• BERT-CCPoem — Pre-trained model for classical Chinese poetry
• Xunzi-LLM — Classical Chinese literature LLM series

---