SAURABH257/gpt2-355M-hinglish-translation-instruct

GPT2-Medium English → Hinglish Translator

A fine-tuned GPT-2 Medium model for translating English text into Hinglish (Hindi written using the Latin alphabet).

Model Overview

This model has been fine-tuned on the findnitai/english-to-hinglish dataset to generate natural Hinglish translations from English input. The model is designed for conversational text, social media content, chatbots, and localization use cases targeting Indian audiences.

Example

Input

How are you doing today?

Output

Aaj tum kaise ho?

---

Model Details

Attribute	Value
Base Model	GPT-2 Medium
Task	English → Hinglish Translation
Dataset	findnitai/english-to-hinglish
Languages	English, Hinglish
License	MIT
Framework	Hugging Face Transformers
Developed By	Saurabh Prajapati

---

Training Dataset

Dataset used:

• findnitai/english-to-hinglish

The dataset contains paired English and Hinglish sentences suitable for supervised fine-tuning.

Example:

{
  "en": "What are you doing?",
  "hi_ng": "Tum kya kar rahe ho?"
}

---

Intended Use

Direct Use

• English to Hinglish translation
• Conversational AI
• Customer support chatbots
• Social media content generation
• WhatsApp and SMS automation

Downstream Applications

• Multilingual assistants
• Marketing content localization
• Hinglish content generation
• Voice assistant backends

---

Out-of-Scope Use

This model is not intended for:

• Legal translation
• Medical translation
• Financial advice
• Safety-critical applications
• Professional human translation replacement

Generated outputs should always be reviewed before production deployment.

---

Quick Start

Installation

pip install transformers torch

---

Training Procedure

Input Format

Training examples were formatted as:

Translate English to Hinglish:

English: How are you?
Hinglish: Tum kaise ho?

Training Configuration

Parameter	Value
Base Model	GPT-2 Medium
Parameters	355M
Memory	1.7Gb
Context_size	1024
Emb_dim	1024
n_layers	24
n_heads	16
Vocab_size	50257
Optimizer	AdamW
Learning Rate	5e-5
Precision	FP16
Framework	Google Colab
Task	Causal Language Modeling

---

Evaluation

Metric

• Accuracy
• 78% Accuracy on the Test Dataset (100 Samples)

Sample Results

English	Hinglish
How many accidents are on the i 5 north freeway?	i 5 north freeway par kitne accidents hue he?
Report the weather in New Zealand to me	mere liye New Zealand ka weather report bataiye
Remind me every month to pay my mortgage payment	mujhe har mahine har mahine mera mortgage payment karne ke liye yaad dilaye

The model performs well on common conversational English sentences and generates natural Hinglish outputs.

---

Limitations

• Hinglish spelling is not standardized.
• Performance may degrade on long paragraphs.
• Technical and domain-specific terminology may not translate well.
• Outputs may occasionally mix English and Hindi terms.

---

Bias and Risks

This model inherits biases present in:

1. GPT-2 pretraining data.
2. The English-Hinglish fine-tuning dataset.

---

Technical Specifications

Architecture

• GPT-2 Medium
• Decoder-only Transformer
• ~355 Million Parameters

Software Stack

• Python
• PyTorch
• Hugging Face Transformers
• Hugging Face Datasets

---

Acknowledgements

• OpenAI GPT-2
• Hugging Face Transformers
• Hugging Face Datasets
• findnitai/english-to-hinglish dataset

---

Citation

@misc{prajapati2026gpt2hinglish,
  author = {Saurabh Prajapati},
  title = {GPT2-Medium English to Hinglish Translator},
  year = {2026},
  publisher = {Hugging Face}
}

---

Contact

For questions, issues, or collaboration opportunities, please open an issue on the Hugging Face repository.