| ## 🕊️ TagoreX – A Bengali Text Generator Inspired by Tagore | |
| **Model name:** `SwastikGuhaRoy/TagoreX` | |
| **Base model:** `GPT-2` with LoRA adapters [(based on `AddaGPT2.0`)](https://huggingface.co/SwastikGuhaRoy/AddaGPT2.0) | |
| **Language:** Bengali | |
| **Author:** Swastik Guha Roy (`@SwastikGuhaRoy`) | |
| **License:** MIT | |
| **Model size:** \~124M parameters | |
| **Trained on:** Curated (but imperfect) corpus of Rabindranath Tagore’s writings | |
| **Intended use:** Poetic and philosophical Bengali text generation | |
| **Demo app:** [TagoreX + Gemini Streamlit App](https://tagorexgemini.streamlit.app) | |
| --- | |
| ### 📘 Model Description | |
| **TagoreX** is a fine-tuned version of `AddaGPT2.0` — a small GPT-2 model adapted for Bengali using LoRA (Low-Rank Adaptation). | |
| This model was trained on literary works of Rabindranath Tagore as a tribute. | |
| The model continues a given Bengali prompt in a Tagore-like poetic tone. It generates \~256 tokens, which are then optionally refined by Gemini AI in a downstream application. | |
| --- | |
| ### 🔧 Technical Details | |
| * **Architecture**: GPT-2 (117M parameters) | |
| * **Training strategy**: Full fine-tuning | |
| * **Epochs**: 22 (symbolically referencing “২২শে শ্রাবণ”) | |
| * **Max sequence length**: 256 tokens | |
| * **Tokenizer**: AutoTokenizer from the base model | |
| * **Framework**: PyTorch + Transformers | |
| --- | |
| ### 📂 Training Data | |
| The dataset includes poems, prose and other works from Rabindranath Tagore which is [publicly available](https://archive.org/details/RABINDRARACHANABALI/). [The dataset can be accessed in a consolidated .txt format from here](https://huggingface.co/datasets/SwastikGuhaRoy/WorksofTagore) | |
| ⚠️ **Note**: The data may and DOES contain: | |
| * Typos, formatting errors | |
| * OCR issues | |
| * Incomplete or duplicated lines | |
| This model is not a scholarly curation, but an experimental artistic rendering. | |
| --- | |
| ### 🎯 Intended Use | |
| **You can use this model to:** | |
| * Experiment with Bengali poetic text generation | |
| * Create creative writing prompts in Bengali | |
| * Explore Indic LLM capabilities in low-resource settings | |
| This model is **not suitable** for: | |
| * Any commercial or sensitive deployment | |
| * Factual or linguistic accuracy tasks | |
| * Scholarly representation of Tagore’s works | |
| --- | |
| ### 💬 How to Prompt | |
| ```python | |
| from transformers import AutoTokenizer, AutoModelForCausalLM | |
| tokenizer = AutoTokenizer.from_pretrained("SwastikGuhaRoy/TagoreX") | |
| model = AutoModelForCausalLM.from_pretrained("SwastikGuhaRoy/TagoreX") | |
| prompt = "তুমি রবে নীরবে" | |
| inputs = tokenizer(prompt, return_tensors="pt") | |
| outputs = model.generate(**inputs, max_new_tokens=256, do_sample=True, temperature=0.7) | |
| print(tokenizer.decode(outputs[0], skip_special_tokens=True)) | |
| ``` | |
| --- | |
| ### 🚫 Limitations & Disclaimer | |
| * Not aligned, filtered, or safety-trained. | |
| * Most outputs may be incoherent, repetitive, or nonsensical. | |
| * This is **not** meant to reproduce or replace Tagore's literary work. | |
| * The generation reflects training data and randomness — not any human author. | |
| --- | |
| ### 🌏 Why It Matters | |
| TagoreX demonstrates how even small-scale, open models can express poetic and cultural essence in Indic languages — using limited compute and a lot of curiosity. | |
| It aims to inspire communities to build **Indic LLMs**, especially in low-resource and rural settings. | |
| > *"AI doesn’t have to be massive. It can be local, soulful, and deeply human."* | |
| --- | |
| --- | |
| ### 📫 Contact | |
| 📧 Email: `swastikguharoy@googlemail.com` | |
| 💬 Feedback, bugs, or nice generations? I'd love to hear from you! | |
| --- |