gemma-2b-english-quotes

Gemma 2B English Quotes

A LoRA fine-tuned version of Google's Gemma 2B language model, trained on an English quotes dataset. This model is designed to generate and continue English quotes in a quote-and-author format.

Model Details

Model Description

• Developed by: Deepak Singh
• Hugging Face: deepak2310
• Model type: Causal language model with LoRA adapters
• Language: English
• License: Gemma license
• Base model: google/gemma-2b
• Fine-tuning method: Supervised Fine-Tuning (SFT) with LoRA
• Model repository: deepak2310/gemma-2b-english-quotes

Model Summary

This model was fine-tuned on an English quotes dataset using the Google Gemma 2B base model. The training uses Low-Rank Adaptation (LoRA), allowing the model to learn from the quotes dataset without updating all the parameters of the base model.

The model is intended for experimentation with language model fine-tuning, quote generation, and understanding how parameter-efficient fine-tuning works.

Uses

Direct Use

This model can be used for:

• Generating English-style quotes.
• Continuing quote prompts.
• Experimenting with LoRA fine-tuning.
• Learning how to deploy a fine-tuned language model.

Example Prompt

Quote: Be who you are and say what you feel,

The model can be prompted to continue the quote or generate similar text.

Out-of-Scope Use

This model should not be used for:

• Generating misleading quotes attributed to real people.
• Presenting generated content as verified historical quotations.
• High-stakes decision-making.
• Producing harmful, abusive, or misleading content.

Biases, Risks, and Limitations

• The model may generate inaccurate or fictional quotes.
• Generated text may be repetitive or grammatically incorrect.
• The model may produce incorrect author attributions.
• Fine-tuning on quotes does not guarantee factual knowledge about the authors.
• The model's performance has not been comprehensively evaluated.
• Outputs should be reviewed before being published or attributed to real people.

How to Get Started

Installation

pip install transformers peft torch

Load the Fine-Tuned Model

This repository contains LoRA adapter weights. Load the base Gemma model together with the adapter.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model = "google/gemma-2b"
adapter_model = "deepak2310/gemma-2b-english-quotes"

tokenizer = AutoTokenizer.from_pretrained(base_model)

model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.float32,
    device_map="auto"
)

model = PeftModel.from_pretrained(
    model,
    adapter_model
)

model.eval()

prompt = "Quote: Be who you are and say what you feel,"

inputs = tokenizer(
    prompt,
    return_tensors="pt"
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=50,
        do_sample=True,
        temperature=0.7,
        top_p=0.9
    )

print(
    tokenizer.decode(
        outputs[0],
        skip_special_tokens=True
    )
)

Training Details

Training Data

The model was fine-tuned on an English quotes dataset containing quote and author fields.

The training examples were formatted as:

Quote: {quote}
Author: {author}

The dataset contains approximately 2,508 training examples.

Training Procedure

The model was trained using:

• Base model: google/gemma-2b
• Fine-tuning method: LoRA
• Training framework: Transformers, PEFT, and TRL
• Trainer: SFTTrainer
• Training objective: Supervised fine-tuning for causal language modeling

Training Hyperparameters

Training Hardware

• GPU: NVIDIA Tesla T4
• Training steps: 100
• Training runtime: Approximately 171 seconds

Evaluation

Evaluation Method

The model was tested with manually selected quote prompts to observe its generated outputs.

Example prompt:

Quote: I'm selfish,

Another prompt:

Quote: Be who you are and say what you feel,

Results

The model successfully generated text continuations for the tested prompts.

However, no formal evaluation metrics such as BLEU, ROUGE, or perplexity on a held-out test set have been calculated.

Therefore, the model's performance should be considered preliminary.

Technical Specifications

Model Architecture

• Architecture: Gemma 2B causal language model
• Fine-tuning: LoRA adapters
• Training objective: Causal language modeling
• Task: Text generation

Software

• Transformers
• PEFT
• TRL
• PyTorch

Framework Versions

• PEFT: 0.20.0

Repository Contents

This repository contains the LoRA adapter weights and tokenizer files.

Typical files include:

• adapter_config.json
• adapter_model.safetensors
• tokenizer.json
• tokenizer_config.json
• training_args.bin
• README.md

Note: This repository contains adapter weights rather than a fully merged Gemma model. The original Gemma 2B base model is required to load the adapter.

License

This model is based on Google's Gemma 2B model and is subject to the Gemma license.

Users should review the original Gemma license and its terms before using or redistributing this model.

Base model: google/gemma-2b

Model Card Authors

Deepak Singh

Hugging Face: deepak2310

Acknowledgements

• Google for the Gemma 2B base model.
• Hugging Face for the Transformers and PEFT ecosystem.
• The TRL library for supervised fine-tuning tools.

Citation

If you use this model in your project, please reference the model repository:

@misc{deepak_gemma_2b_english_quotes,
  author = {Deepak Singh},
  title = {Gemma 2B English Quotes},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {
    https://huggingface.co/deepak2310/gemma-2b-english-quotes
  }
}