p988744/eland-sentiment-zh

Eland Sentiment - Chinese Multi-Domain Sentiment Analysis

A Chinese multi-domain sentiment analysis model fine-tuned on Qwen3-4B using LoRA, achieving 92.03% on financial and 86.85% on multi-domain text.

Model Description

This is a LoRA adapter for Qwen/Qwen3-4B, fine-tuned for Chinese sentiment analysis across multiple domains:

• Financial - Stock market, investment, economic news
• Product - Product reviews, shopping discussions
• Brand - Brand image, corporate reputation
• Organization - Company news, workplace discussions
• Social - Social media, public affairs

Supported Tasks

• Overall Sentiment - Classify the overall sentiment (正面/負面/中立)
• Entity Sentiment - Classify sentiment towards specific entities
• Opinion Sentiment - Classify sentiment of specific opinions
• Stance Detection - Determine if text agrees with an opinion

Performance

Financial Domain

Metric	Score
Macro Average	92.03%
Overall Sentiment	95.00%
Entity Sentiment	93.10%
Opinion Sentiment	85.00%
Agrees with Text	95.00%

Multi-Domain (Product, Brand, Organization, Social)

Metric	Score
Macro Average	86.85%
Overall Sentiment	71.00%
Entity Sentiment	78.57%
Opinion Sentiment	97.83%
Agrees with Text	100.00%

Usage

With Transformers + PEFT

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = "Qwen/Qwen3-4B"
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

# Load LoRA adapter
model = PeftModel.from_pretrained(model, "p988744/eland-sentiment-zh")
tokenizer = AutoTokenizer.from_pretrained("p988744/eland-sentiment-zh")

# Example: Overall sentiment analysis
messages = [
    {"role": "system", "content": "你是一個專業的金融文本情感分析助手。請分析以下文本的整體情感，回答「正面」、「負面」或「中立」。"},
    {"role": "user", "content": "台積電今日股價大漲，市場看好AI需求持續成長。"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=10, do_sample=False)
response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
print(response)  # Expected: 正面

Task Prompts

Overall Sentiment:

System: 你是一個專業的金融文本情感分析助手。請分析以下文本的整體情感，回答「正面」、「負面」或「中立」。
User: [your text]

Entity Sentiment:

System: 你是一個專業的金融文本情感分析助手。請分析以下文本中對「{entity}」的情感，回答「正面」、「負面」或「中立」。
User: [your text]

Opinion Sentiment:

System: 你是一個專業的金融文本情感分析助手。請判斷以下觀點的情感傾向，回答「正面」、「負面」或「中立」。
User: 文本：[text]
觀點：[opinion]

Model Variants

Version	Repository	Use Case
LoRA Adapter	p988744/eland-sentiment-zh	HuggingFace + PEFT
GGUF	p988744/eland-sentiment-zh-gguf	Ollama / llama.cpp
Full Merged	p988744/eland-sentiment-zh-vllm	vLLM

Training Details

Parameter	Value
Base Model	Qwen/Qwen3-4B
Method	LoRA (PEFT)
LoRA Rank	32
LoRA Alpha	64
Trainable Params	66M (1.62%)
Epochs	8
Learning Rate	1e-5
Batch Size	8 (effective)
Training Time	~47 minutes
Hardware	NVIDIA L40S

Dataset

Trained on combined dataset:

• Financial: 1,887 training samples (Taiwan stock market forum and news)
• Multi-domain: 600 training samples (product, brand, organization, social)
• Total: 2,487 training samples

Task distribution:

• Overall sentiment: ~40%
• Entity sentiment: ~30%
• Opinion sentiment: ~30%

Limitations

• Language: Chinese (Traditional) only
• Best Domain: Financial text (92% accuracy)
• Other Domains: Product, brand, organization, social (~87% accuracy)

Citation

@misc{eland-sentiment-zh,
  author = {Eland AI},
  title = {Eland Sentiment: Chinese Financial Sentiment Analysis Model},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/p988744/eland-sentiment-zh}
}

License

Apache 2.0