Content Moderation of Surveillance Search Queries Using Fine-Tuned Generative LLMs
(2025) In Master's Theses in Mathematical Sciences FMSM01 20251Mathematical Statistics
- Abstract (Swedish)
- We study how small, fine-tuned generative large language models (LLMs) can moderate free-text search queries for surveillance video systems. Four open models, Llama 3.2 1B, Llama 3.2 3B, Qwen 2.5 0.5B, and 1.5 B, are trained on six subtasks: safety judgement, problem detection, target detection, span detection, safety explanation, and rephrase. The training combines a public toxicity set with about 785 000 synthetic examples that cover EU AI Act problem classes.
On a balanced synthetic test set, the generative models match or beat a RoBERTa-Base classifier in safety judgement and show significant gains in the harder problem- and target-detection subtasks; Llama 3.2 1 B lifts macro F1 from 0.56 to 0.72 on target detection. They also show... (More) - We study how small, fine-tuned generative large language models (LLMs) can moderate free-text search queries for surveillance video systems. Four open models, Llama 3.2 1B, Llama 3.2 3B, Qwen 2.5 0.5B, and 1.5 B, are trained on six subtasks: safety judgement, problem detection, target detection, span detection, safety explanation, and rephrase. The training combines a public toxicity set with about 785 000 synthetic examples that cover EU AI Act problem classes.
On a balanced synthetic test set, the generative models match or beat a RoBERTa-Base classifier in safety judgement and show significant gains in the harder problem- and target-detection subtasks; Llama 3.2 1 B lifts macro F1 from 0.56 to 0.72 on target detection. They also show improved performance on span detection and produce short reasons and rephrases, features the classifier cannot offer. On real-world, publicly available, albeit imbalanced data, the results were mixed but still favored the generative LLMs for the more nuanced tasks.
Although the generative models need more memory and inference time, their high performance, richer output, and simple prompt interface make them a promising core for an open, transparent moderation pipeline in video search. Key open issues are speed, subtask consistency, and resistance to adversarial text. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9200534
- author
- Bakly, Ali LU and Than, Davy
- supervisor
- organization
- course
- FMSM01 20251
- year
- 2025
- type
- H2 - Master's Degree (Two Years)
- subject
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUTFMS-3528-2025
- ISSN
- 1404-6342
- other publication id
- 2025:E71
- language
- English
- id
- 9200534
- date added to LUP
- 2025-06-18 12:55:10
- date last changed
- 2025-06-26 14:38:04
@misc{9200534, abstract = {{We study how small, fine-tuned generative large language models (LLMs) can moderate free-text search queries for surveillance video systems. Four open models, Llama 3.2 1B, Llama 3.2 3B, Qwen 2.5 0.5B, and 1.5 B, are trained on six subtasks: safety judgement, problem detection, target detection, span detection, safety explanation, and rephrase. The training combines a public toxicity set with about 785 000 synthetic examples that cover EU AI Act problem classes. On a balanced synthetic test set, the generative models match or beat a RoBERTa-Base classifier in safety judgement and show significant gains in the harder problem- and target-detection subtasks; Llama 3.2 1 B lifts macro F1 from 0.56 to 0.72 on target detection. They also show improved performance on span detection and produce short reasons and rephrases, features the classifier cannot offer. On real-world, publicly available, albeit imbalanced data, the results were mixed but still favored the generative LLMs for the more nuanced tasks. Although the generative models need more memory and inference time, their high performance, richer output, and simple prompt interface make them a promising core for an open, transparent moderation pipeline in video search. Key open issues are speed, subtask consistency, and resistance to adversarial text.}}, author = {{Bakly, Ali and Than, Davy}}, issn = {{1404-6342}}, language = {{eng}}, note = {{Student Paper}}, series = {{Master's Theses in Mathematical Sciences}}, title = {{Content Moderation of Surveillance Search Queries Using Fine-Tuned Generative LLMs}}, year = {{2025}}, }