abhiprd20/Maithili_Sentiment_8K

🗣️ Maithili Sentiment Dataset

55,000 sentences for 3-class sentiment analysis in Maithili (मैथिली) — Devanagari script

---

📌 Dataset Description

Maithili Sentiment Dataset is a large-scale, richly annotated sentiment analysis resource for Maithili (मैथिली) — one of the 22 scheduled languages of the Indian Constitution, spoken by approximately 34 million people across the Mithila region of Bihar, India and the Terai region of Nepal.

Despite its large speaker base and official constitutional recognition since 2003, Maithili has been almost entirely absent from NLP research. Standard multilingual benchmarks contain near-zero Maithili representation, and even large pretrained models like mBERT and XLM-RoBERTa have minimal Maithili coverage in their pretraining corpora.

This dataset is among the first large-scale sentiment resources for Maithili, containing 55,000 sentences annotated across three classes:

• 😞 Negative — text expressing dissatisfaction, criticism, or distress
• 😐 Neutral — text that is factual, descriptive, or indifferent
• 😊 Positive — text expressing satisfaction, happiness, or praise

Each entry includes rich metadata covering domain, region, speaker type, sentiment intensity, and code-switching detection — making it suitable for sentiment classification, sociolinguistic analysis, dialect variation research, and cross-lingual transfer studies.

---

📊 Dataset Statistics

Property	Value
Total rows	55,000
Language	Maithili (mai)
Script	Devanagari
Classes	3 (negative, neutral, positive)
Label balance	~18,333 per class
Includes English translation	✅ Yes
Includes sentiment intensity	✅ Yes
Includes code-switch flag	✅ Yes

---

⚡ Quick Start

from datasets import load_dataset

dataset = load_dataset("abhiprd20/Maithili_Sentiment_8K")
train   = dataset["train"]

print(f"Total rows: {len(train)}")
print(train[0])

---

🔍 More Examples

from datasets import load_dataset

dataset  = load_dataset("abhiprd20/Maithili_Sentiment_8K")
train    = dataset["train"]
ID2LABEL = {0: "negative", 1: "neutral", 2: "positive"}

# Filter to verified rows only (recommended for training)
verified = train.filter(lambda x: x["verified"] == True)
print(f"Verified rows: {len(verified)}")

# Print a sample from each class
for label_id in [0, 1, 2]:
    sample = train.filter(lambda x: x["label"] == label_id)[0]
    print(f"\n[{ID2LABEL[label_id].upper()}]")
    print(f"  Maithili : {sample['text']}")
    print(f"  English  : {sample['english_translation']}")

Expected output:

[NEGATIVE]
  Maithili : हमरा ई खाना नीक नहि लागल।
  English  : I did not like this food.

[NEUTRAL]
  Maithili : हम काल्हि पटना जाइब।
  English  : I will go to Patna tomorrow.

[POSITIVE]
  Maithili : ई पोथी बहुत नीक ऐछ।
  English  : This book is very good.

---

🧪 Fine-Tuning a Classifier

from datasets import load_dataset
from transformers import (
    AutoTokenizer,
    AutoModelForSequenceClassification,
    Trainer,
    TrainingArguments
)

dataset   = load_dataset("abhiprd20/Maithili_Sentiment_8K")
tokenizer = AutoTokenizer.from_pretrained("bert-base-multilingual-cased")

# Use verified rows for cleanest training signal
train_data = dataset["train"].filter(lambda x: x["verified"] == True)

def tokenize(batch):
    return tokenizer(
        batch["text"],
        truncation=True,
        padding="max_length",
        max_length=128
    )

train_tok = train_data.map(tokenize, batched=True)
train_tok = train_tok.rename_column("label", "labels")
train_tok.set_format("torch", columns=["input_ids", "attention_mask", "labels"])

model = AutoModelForSequenceClassification.from_pretrained(
    "bert-base-multilingual-cased",
    num_labels=3,
    id2label={0: "negative", 1: "neutral", 2: "positive"},
    label2id={"negative": 0, "neutral": 1, "positive": 2}
)

⚠️ Important: Always use bert-base-multilingual-cased or xlm-roberta-base for Maithili. bert-base-uncased does not support Devanagari script and will produce random results.

---

🗂️ Dataset Structure

Column	Type	Description
id	int	Unique row identifier
text	string	Maithili sentence in Devanagari script
label	int	Sentiment class: 0 = negative, 1 = neutral, 2 = positive
label_int	int	Integer copy of label (same values, retained for compatibility)
language	string	Language identifier (Maithili)
domain	string	Broad topic domain (e.g. social, news, everyday life)
sub_domain	string	Finer-grained topic category
source	string	Origin of the text
speaker_type	string	Type of speaker or author
region	string	Geographic region of speaker (Bihar / Nepal Terai / other)
script	string	Script used (Devanagari)
verified	bool	Whether the entry passed human quality verification
english_translation	string	English translation of the Maithili sentence
sentiment_intensity	string	Strength of sentiment (weak / moderate / strong)
contains_code_switch	bool	Whether the sentence mixes Maithili with Hindi or English

---

🏷️ Label Mapping

Label ID	Label	Meaning
0	negative	Dissatisfaction, criticism, distress, anger
1	neutral	Factual, descriptive, balanced, indifferent
2	positive	Satisfaction, happiness, praise, appreciation

---

🌍 Why Maithili?

Maithili is spoken by ~34 million people, yet has virtually no NLP infrastructure. It is:

• One of the 22 scheduled languages of the Indian Constitution
• Officially recognised since 2003 under the Eighth Schedule
• Spoken across Bihar, Jharkhand, and the Terai region of Nepal
• Written in both Devanagari and the classical Tirhuta script

A cross-lingual transfer study using this dataset found that English BERT collapses from 84.58% on English to 33.33% on Maithili in zero-shot evaluation — a 51.25% accuracy drop — confirming the severity of the resource gap. Fine-tuning XLM-RoBERTa on a subset of this dataset recovers performance to 85.63%, showing that even modest Maithili data significantly closes the gap.

---

🔗 Related Resources

Resource	Link
Maithili sentiment model (trained on this dataset)	abhiprd20/maithili-sentiment-model
Bhojpuri sentiment model	abhiprd20/bhojpuri-sentiment-model
Hindi sentiment model	abhiprd20/hindi-sentiment-model
English baseline model	abhiprd20/nlp-sentiment-model

---

⚖️ License

This dataset is released under the Apache License 2.0 research and commercial use.

Copyright 2026 Abhimanyu Prasad

---

📎 Citation

If you use this dataset in your research or project, please cite and contact through hugging face or gmail:

@dataset{prasad2026maithili,
  author       = {Prasad, Abhimanyu},
  title        = {Maithili Sentiment Dataset: 55,000 Sentences for 3-Class Sentiment Analysis},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/abhiprd20/Maithili_Sentiment_8K}},
  note         = {55,000 Maithili sentences in Devanagari script with 3-class sentiment labels,
                  domain metadata, and English translations}
}

---

👤 Author

Abhimanyu Prasad 🤗 Hugging Face: abhiprd20 E-mail: [8292aniarc@gmail.com] 🤖 Model: abhiprd20/maithili-sentiment-model

---

If this dataset helped your research, consider giving it a ⭐ — it helps others working on low-resource Indic NLP find it too!