v1.1.0

March 30, 2025:

• Added round 2 experiment results in results_round2, electra_finetune, and gpt_finetune_experiments_round2.ipynb
• Corrected round one E7-G4O-ELFT data in results (was E14 now E19 in latest numbering), gpt_finetune_experiments.ipynb and statistics.ipynb
• Updated research_paper.pdf, which addresses some reviewer feedback and includes 2 rounds of experiments
• Updated requirements.txt

v1.0.0

Sentiment 1.0.0 Release

Initial release of project codebase and research for "ELECTRA and GPT-4o: Cost-Effective Partners for Sentiment Analysis".

Added

• Data: Merged dataset is published in data directory, with links on README to Hugging Face dataset. Data processing to create the Merged dataset is shown in data_processing.ipynb.
• Classifier Model: PyTorch code with DDP support for BERT-based encoder models with custom pooling and either a classifier head or fully fine-tuned is in classifier.py
• Classifier Fine-Tuning Code: Interactive fine-tuning finetune.py program using PyTorch with DDP support that can be used to train a classifier head or fine-tune any number of layers on BERT, RoBERTa, ELECTRA, etc.
• Classifier Fine-Tune Logs: Logs of ELECTRA Base/Large baselines (classifier head only) and full fine-tunes (all layers) are in electra_finetune
• Fine-Tuned ELECTRA Models: Links on README to fine-tuned ELECTRA models on Hugging Face
• GPT Fine-Tuning Code: Code to process data into JSONL and use OpenAI API to fine-tune GPT models is shared in gpt_finetune_experiments.ipynb
• Experiment Runs: All the research experiments involving GPT are documented in gpt_finetune_experiments.ipynb.
• Experiment Results: The predictions and metrics of each experimental run is published in results. Statistics for hypotheses evaluation are in statistics.
• Research Paper: A PDF of the research paper is included, as well as links to the preprint published on arXiv.