p988744/eland-sentiment-zh

Eland Sentiment Chinese Financial Dataset

A Chinese financial sentiment analysis dataset for Taiwan stock market text.

Dataset Configs

This dataset provides two configurations for different use cases:

Config	Format	Use Case
raw	Flat JSON (text, overall, ...)	Data analysis, custom preprocessing
chat	Messages array (messages[])	LLM fine-tuning (SFT)

Quick Start

from datasets import load_dataset

# Load RAW format (for analysis)
raw_dataset = load_dataset("p988744/eland-sentiment-zh", "raw")

# Load CHAT format (for training)
chat_dataset = load_dataset("p988744/eland-sentiment-zh", "chat")

Dataset Statistics

RAW Config

Split	Samples
train	999
validation	300
test	300
Total	1,599

CHAT Config

Split	Samples
train	1,887
validation	360
test	360
Total	2,607

Sentiment Analysis Tasks

Both configs support three analysis tasks:

1. Overall Sentiment - Classify the overall sentiment of the text
2. Entity Sentiment - Classify sentiment towards specific entities (companies, products)
3. Opinion Sentiment - Classify sentiment of specific opinions in text

Labels

• 正面 (Positive)
• 中立 (Neutral)
• 負面 (Negative)

Data Format

RAW Config

{
  "text": "台積電今日股價大漲...",
  "overall": "正面",
  "task": "overall",
  "source": "forum_01"
}

With entity:

{
  "text": "台積電今日股價大漲...",
  "overall": "正面",
  "entity": "台積電",
  "entity_sentiment": "正面",
  "task": "entity"
}

CHAT Config

{
  "messages": [
    {"role": "system", "content": "你是一個專業的情緒分析助手..."},
    {"role": "user", "content": "分析以下文本的情緒。\n\n文本：台積電今日股價大漲..."},
    {"role": "assistant", "content": "{\"overall\": \"正面\"}"}
  ]
}

Token Statistics

RAW Config

Split	Samples	Estimated Tokens
train	999	~146K
validation	300	~44K
test	300	~44K
Total	1,599	~234K

CHAT Config

Split	Samples	Estimated Tokens
train	1,887	~380K
validation	360	~73K
test	360	~73K
Total	2,607	~526K

Token counts estimated using Qwen3-4B tokenizer.

Calculate Tokens

from datasets import load_dataset
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")

# RAW config
raw_dataset = load_dataset("p988744/eland-sentiment-zh", "raw")
raw_tokens = sum(
    len(tokenizer.encode(s["text"], add_special_tokens=False))
    for s in raw_dataset["train"]
)
print(f"RAW train tokens: {raw_tokens:,}")

# CHAT config
chat_dataset = load_dataset("p988744/eland-sentiment-zh", "chat")
chat_tokens = sum(
    sum(len(tokenizer.encode(m["content"], add_special_tokens=False))
        for m in s["messages"])
    for s in chat_dataset["train"]
)
print(f"CHAT train tokens: {chat_tokens:,}")

Usage Examples

For Data Analysis (RAW)

from datasets import load_dataset

dataset = load_dataset("p988744/eland-sentiment-zh", "raw")

# Filter by task type
overall_samples = dataset["train"].filter(lambda x: x["task"] == "overall")
entity_samples = dataset["train"].filter(lambda x: x["task"] == "entity")

# Analyze sentiment distribution
from collections import Counter
sentiments = Counter(s["overall"] for s in dataset["train"])
print(sentiments)

For LLM Training (CHAT)

from datasets import load_dataset
from trl import SFTTrainer

dataset = load_dataset("p988744/eland-sentiment-zh", "chat")

# Ready for SFTTrainer
trainer = SFTTrainer(
    model=model,
    train_dataset=dataset["train"],
    # ... other configs
)

Associated Model

This dataset was used to train p988744/eland-sentiment-zh, which achieves 89.38% Reliability on the RGL benchmark.

Related Datasets

Dataset	Task
eland-stance-zh-data	Stance Detection
eland-entity-sentiment-zh-data	Entity Sentiment

License

Apache 2.0

Citation

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