youtube-sentiment-model-telugu

Model Overview

This model is a fine-tuned version of AmaanP314/youtube-xlm-roberta-base-sentiment-multilingual. While the base model was trained on multilingual English-only YouTube comments, this version has been fine-tuned on a large dataset of Telugu comments, enabling it to classify Telugu (native script), transliterated Telugu, and English YouTube comments into sentiment categories: Negative, Neutral, and Positive.

Model Details

Base model: AmaanP314/youtube-xlm-roberta-base-sentiment-multilingual

Fine-tuned for: Telugu + English YouTube comment sentiment analysis

Languages Supported:

Telugu (native script)

Transliterated Telugu

English

Labels:

0: Negative

1: Neutral

2: Positive

Dataset & Labeling

Source: Comments were extracted from YouTube using the YouTube Data API.

Comment Count:

Train set: 73,943 comments

Validation set: 8,216 comments

Labeling Method: Comments were labeled using Gemini 1.5 Pro (Google’s LLM) via a sentiment classification prompt to auto-assign one of the three sentiment classes.

How to Use

The model can be used via an API endpoint or loaded locally using the Hugging Face Transformers library. For example, using Python:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "gajula21/youtube-sentiment-model-telugu"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

comments = [
    "ఈ సినిమా చాలా బాగుంది!",  
    "ఈ వీడియో చాలా బోరు పడింది",  
    "ఇది మామూలు వీడియో",  
]

inputs = tokenizer(comments, return_tensors="pt", padding=True, truncation=True)

with torch.no_grad():
    outputs = model(**inputs)

predictions = torch.argmax(outputs.logits, dim=1)
label_mapping = {0: "Negative", 1: "Neutral", 2: "Positive"}
sentiments = [label_mapping[p.item()] for p in predictions]

print(sentiments)

Training Configuration

Framework: Hugging Face Transformers (PyTorch)

Tokenizer: AutoTokenizer from base model

Loss Function: CrossEntropyLoss with label_smoothing=0.1

Batch Size: 1176 (per device)

Gradient Accumulation Steps: 2

Learning Rate: 1e-5

Weight Decay: 0.05

Epochs: 3

Evaluation Strategy: Every 125 steps

Early Stopping: Patience of 5 evaluation steps

Mixed Precision: Enabled (fp16)

Evaluation Results

Citation [optional]

@misc{gajula21_youtube_sentiment_2025,
  author = {Gajula Vivek},
  title = {Telugu-English YouTube Sentiment Classifier},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/gajula21/youtube-sentiment-model-telugu}},
}