Python Project 1 - LLM Evaluation

Evaluating LLM Truthfulness with Prompting, Fine-Tuning and RAG

This project evaluates how different large language model adaptation strategies affect the truthfulness and semantic quality of generated answers.

Five variants of Qwen 2.5 are compared using TruthfulQA-derived question and reference-answer data:

1. Qwen 2.5 baseline
2. Baseline + system prompt
3. LoRA fine-tuned Qwen 2.5
4. Baseline + Retrieval-Augmented Generation (RAG)
5. Qwen 2.5 Instruct + RAG

The repository contains the model-output datasets required to reproduce the evaluation.

Research Question

How do prompt engineering, LoRA fine-tuning, and retrieval-augmented generation affect the truthfulness and semantic similarity of Qwen 2.5 responses?

Data

The repository distinguishes between the source benchmark data and the generated model outputs.

Source benchmark

data/TruthfulQA-2.csv

This is the TruthfulQA dataset used in the project. It contains 817 benchmark questions, including question type, category, correct reference answers, incorrect reference answers, and source information.

Generated model outputs

File	Model
qwen_qa_results_no_system_prompt.csv	Qwen 2.5 baseline
qwen_qa_results_system_prompt.csv	Baseline + system prompt
qwen_qa_results.csv	LoRA fine-tuned model
qwen_base_rag_answers.csv	Baseline + RAG
qwen_instruct_rag_answers.csv	Qwen Instruct + RAG

analysis.py merges each generated-answer file back to the canonical TruthfulQA benchmark by question before computing the evaluation metrics.

Evaluation

The analysis compares each generated response with the available correct and incorrect reference answers.

For each response:

1. BLEURT is used to identify the best-matching reference answer.
2. BLEU and ROUGE measure lexical overlap with the selected reference.
3. Sentence-transformer embeddings are used to calculate cosine similarity.
4. The generated response is classified as reference-based correct when its best-matching reference belongs to the correct-answer set.
5. Confidence-filtered correctness is calculated for responses with cosine similarity above a predefined threshold.

The default threshold used in analysis.py is:

CONFIDENCE_THRESHOLD = 0.60

Original Reported Results

The original project results are included in the results/ folder, so the repository documents both the input data and the reported outputs without requiring the models to be rerun.

Model	High-confidence correctness	Mean cosine similarity	High-confidence responses
Baseline	42.45%	0.489	318
Baseline + system prompt	44.98%	0.694	618
LoRA fine-tuned	60.27%	0.577	516
Baseline + RAG	47.60%	0.413	292
Qwen Instruct + RAG	49.64%	0.729	685

The original notebook also reports separate results for adversarial and non-adversarial questions; these are included as CSV files in results/.

Reproducibility

Create a Python environment and install the dependencies:

pip install -r requirements.txt

Then run:

python analysis.py

The script reads all five files from data/ and writes the evaluation outputs to results/.

Repository Structure

Python Project 1 - LLM Evaluation/
├── README.md
├── analysis.py
├── requirements.txt
├── data/
│   ├── TruthfulQA-2.csv
│   ├── qwen_qa_results_no_system_prompt.csv
│   ├── qwen_qa_results_system_prompt.csv
│   ├── qwen_qa_results.csv
│   ├── qwen_base_rag_answers.csv
│   └── qwen_instruct_rag_answers.csv
└── results/
    ├── original_total_metrics.csv
    ├── original_adversarial_metrics.csv
    ├── original_non_adversarial_metrics.csv
    ├── validation_report.csv

Validation

The packaged repository was checked for:

• Python syntax validity of analysis.py;
• presence of the TruthfulQA source dataset;
• presence of all five model-output files;
• one-to-one matching of all 817 questions between each model-output file and the benchmark.

The full metric recomputation was not executed in this packaging environment because it requires external runtime dependencies and model downloads (evaluate, BLEURT, and sentence-transformers). These dependencies are listed in requirements.txt.

Methods and Python Packages

The analysis uses:

• pandas
• numpy
• evaluate
• sentence-transformers
• scikit-learn
• BLEURT
• BLEU
• ROUGE

Skills Demonstrated

This project demonstrates experience with:

• Large language model evaluation
• Natural language processing
• Qwen 2.5
• LoRA fine-tuning
• Retrieval-Augmented Generation
• Prompt engineering
• Transformer embeddings
• Semantic similarity
• NLP evaluation metrics
• Comparative model evaluation
• Reproducible Python workflows

Contributors

• Laura Maria Fetz
• Martin Turna
• Bart Amin

References

• Lin, S., Hilton, J., & Evans, O. (2021). TruthfulQA: Measuring How Models Mimic Human Falsehoods.
• Papineni, K., Roukos, S., Ward, T., & Zhu, W. (2002). BLEU: A Method for Automatic Evaluation of Machine Translation.
• Lin, C. (2004). ROUGE: A Package for Automatic Evaluation of Summaries.
• Sellam, T., Das, D., & Parikh, A. (2020). BLEURT: Learning Robust Metrics for Text Generation.