Sentiment Analysis with LLMs: Evaluating QLoRA Fine-Tuning, Instruction Strategies, and Prompt Sensitivity
Bibliographic record
Abstract
In this study, we explore the performance of large language models (LLMs) and provide insights into the strengths and weaknesses of these models in sentiment analysis. In addition, we explore the impact of various instruction methods and fine-tuning techniques on the models' performance and analyze how sensitive they are to different prompts. Although there are studies comparing LLMs on sentiment analysis tasks, there is a limited number of studies on fine-tuning strategies, instruction variety, and prompt sensitivity of LLMs for sentiment analysis. For this purpose, we used Llama2-7B, Lama3-8B, Mistral-7B, Gemma-7B, and Bloom-7B and compared them with BERT and Roberta. We evaluated fine-tuned and original chat versions with zero and few-shot experiments. We generated various instruction strategies and observed their effects on the performance of the models. We also compared Quantized Low-Rank Adapters (QLoRA) fine-tuning and full fine-tuning with all parameters. The fine-tuned versions performed better compared to the original chat versions on almost all tasks. The fine-tuned and chat versions of Llama3-8B and Mistral-7B outperformed others on these tasks, while the Llama2-7B-chat model often produced invalid responses. Our research indicates that QLoRA is a practical substitute for full fine-tuning, as it lessens computational demands without sacrificing performance. Fine-tuned models were less affected by variations in test prompts compared to chat models. This is likely because their specialized training allows them to better handle queries within their domain. In this exploratory study, we provide insights into the potential of LLMs in sentiment analysis tasks.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.002 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.002 | 0.001 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".