#Important:

This repo is split into two projects, one of which focuses on Sentiment Analysis of users' product reviews, and the second one which focuses on trying to fine-tune a local LLM to impersonate a person's writing style.

SentimentAnalysisCUDA

Running envonment:

    source .venv/bin/activate
    uv pip install -r requirements.txt

Loading Dataset

python src/download_data.py

Running Training

python -m src.train --model transformer --run-name transformer_run_01

--model [transformer|lstm]
--run-name [any name for the run, e.g. transformer_run_01]
--num-heads [number of heads for transformer, e.g. 4]
--num-layers [number of layers for transformer, e.g. 2]
--ff-dim [feedforward dimension for transformer, e.g. 128]

TrumpTweetGenerationCUDA

Colab ver.

Fine-tuning DistilGPT-2 on Trump's Legacy dataset to generate Trump-style tweets using CUDA acceleration.

Running environment (Google Colab):

No local setup required. Simply open the notebook in Google Colab and run the first cell to install the remaining required packages:

!pip install -q torch transformers datasets pandas kaggle

## Informations & parameters
--notebook [Trump_Tweet_DistilGPT2_FineTuning.ipynb]
--model [distilgpt2]
--epochs [number of training epochs, default: 3]
--batch-size [batch size for training, default: 16]
--lr [learning rate, default: 5e-5]

Running locally:

Install the required packages using pip. If you are installing globally outside of a virtual environment, you may need the --break-system-packages flag:

pip install -r requirements_server.txt

The training script automatically downloads the Trump's Legacy dataset from Kaggle. To enable this, you need to set up your Kaggle API credentials. Place your kaggle.json (downloaded from Kaggle -> Settings -> Create New Token) into the .kaggle directory:

mkdir -p ~/.kaggle
echo '{"username":"YOUR_KAGGLE_USERNAME","key":"YOUR_KAGGLE_KEY"}' > ~/.kaggle/kaggle.json
chmod 600 ~/.kaggle/kaggle.json

To start the fine-tuning process, run:

python3 scripts/train_qwen_lora.py

# low VRAM optimization (Optional)
# The default configuration is optimized for larger GPUs. If you have limited VRAM (e.g.8 GB), to prevent Out of Memory (OOM) errors use:

python3 scripts/train_qwen_lora.py --batch-size 2 --grad-accum 16 --no-compile

Inference & Text Generation

Once the model is trained and the LoRA adapter is saved, you can use the generation script to test how the fine-tuned model performs and verify the results. The script loads the base model, applies your custom adapter, and generates text based on a given prompt.

Usage

To run the text generation with the default prompt ("The fake news"), use the following command:

python3 scripts/Ccheck_results.py

You can pass any custom starting phrase using the --prompt flag to see how the fine-tuned model completes the sentence in Trump's characteristic style:

python3 scripts/generate.py --prompt "China is"