halu_control

Methods for controlling hallucinations for LLM in Summarization

Blog post: https://vectara.com/blog/reducing-hallucinations-in-llms/

Benchmark settings

MODEL	Strategy	Consistency Rate	Answer Rate	Average Length
Mistral-7B-Instruct-v0.1	Greedy	93.2	100.0	93.5
Mistral-7B-Instruct-v0.1	num_beam = 10	95.3	100.0	127.7
Mistral-7B-Instruct-v0.1	Greedy + DoLA	93.7	100.0	93.6
Mistral-7B-Instruct-v0.1	Greedy + DPO(LoRA)	95.8	100.0	97.0
Mistral-7B-Instruct-v0.1	Greedy + Fava	93.7	100.0	93.3
Mistral-7B-Instruct-v0.1	DPO(LoRA) + num_beam=10	96.9	100.0	123.7
Mistral-7B-Instruct-v0.1	Best_of_N + Temperature=0.7 + n=10	99.3	100.0	89.6

Note: Prompt slightly different from the orginal HHEM benchmark, causing different numbers.

How to reproduce the experiments

1. Download the leaderboard dataset (https://huggingface.co/spaces/vectara/leaderboard/raw/main/src/datasets/leaderboard_dataset.csv)
2. Generate the model response generated.csv, see methods below
3. Run evaluation on the reposnse file

python -c "from leaderboard import run_eval;run_eval('generated.csv')"

Methods

Baselines

1. Greedy/Beam Search

• Notebook: 1_decoding.ipynb

2. Best of N sampling

• Notebook: 5_best_of_n.ipynb

DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models

• Paper Link
• Notebook: 2_dola.ipynb

Fine-tuning Language Models for Factuality

• Paper Link
• Notebook: 3_dpo.ipynb
• Training code: dpo_training.py
• Note: our setting is different from the original paper, we used CNN/Dailymail+XSum+VitaminC as the source dataset and HHEM model as the reference metric for factuality.

Fine-grained Hallucination Detection and Editing For Language Models

• Web Link
• Notebook: 4_fava.ipynb