In contemporary online interactions, identifying inappropriate language and safeguarding minors from harmful communication is a critical challenge. This study explores the use of Large Language Models (LLMs) to analyze text, detecting patterns indicative of age-specific language and the presence of sexual or pornographic references. A fine-tuning of the LLaMAntino model was performed, using a dataset of synthetically generated sentences designed to replicate realworld scenarios. The fine-tuned model demonstrated enhanced performance compared to its baseline (given by LLaMAntino 3 ANITA 8B), providing detailed and context-sensitive explanations for its classifications. The results highlight the potential of LLMs in addressing sensitive linguistic phenomena with precision, offering a foundation for detecting indirect combinations of sexual references in conversations involving minors. Future work can focus on incorporating real conversational data and involving subject matter experts to refine the model's interpretability and reliability. Additionally, the exploration of advanced architectures and fine-tuning techniques will be considered to further balance model complexity and processing efficiency. 1 In the modern digital age, the prevalence of harmful communication in online interactions has become an increasingly pressing issue, particularly when it comes to protecting minors. Inappropriate language, which may include age-improper expressions or sexually explicit content, poses significant risks to the emotional and psychological well-being of young individuals [1]. If left unaddressed, these dangerous interactions can contribute to unsafe digital environments, foster exploitative behaviours, and erode trust in online platforms. Identifying and mitigating such communication is crucial to ensuring a safer and more supportive space for minors in the digital landscape. Modern Large Language Models (LLMs) offer a promising solution to this challenge due to their ability to analyze and interpret complex textual patterns and contexts. These advanced models,
when fine-tuned with targeted datasets, can effectively identify sensitive linguistic phenomena and provide detailed, context-aware explanations. By leveraging the power of LLMs, researchers and developers can create tools to detect harmful communication patterns, safeguard vulnerable populations, and promote healthier digital interactions.
Recent research on large language models (LLMs) in the field of harmful language detection
and safeguarding minors has seen notable advancements. Studies have explored techniques for
identifying harmful communication patterns, including predatory behavior, sexually explicit
content, and age-inappropriate language. This involves fine-tuning models with
domainspecific datasets and leveraging transfer learning to improve their ability to detect subtle
linguistic cues indicative of harmful interactions, as demonstrated in [
Recent work has also focused on improving LLMs' capabilities to process multimodal inputs
and apply advanced reasoning techniques for better contextual understanding. For instance,
studies such as [
In general, these studies collectively underscore the growing potential of LLMs to address the
challenges of harmful communication in digital spaces, particularly in contexts involving
minors. By fine-tuning models on specialized datasets, incorporating multimodal reasoning, and
addressing ethical concerns, researchers are laying a foundation for safer and more responsible
AI applications in safeguarding vulnerable populations [
Current LLMs face challenges in accurately determining and explaining whether a text contains age-specific language and/or sexual (abuse) references. To address this, we focused on finetuning a single model to specialize in these tasks more effectively. Using LLaMAntino as our baseline, we examined its ability to provide detailed explanations for requests like: "Analyses the following text and indicates whether the language used belongs to a minor or an adult, explaining:" and " Analyses the following text and indicates whether the language used contains sexual references or porn, explaining:".
The initial responses from the model were not sufficiently detailed, leading us to perform finetuning using a carefully constructed dataset. We annotated 500 synthetic samples to train and test the model, ensuring the dataset covered a variety of cases involving minor/adult language distinctions and adult sexual references. These synthetic samples were generated using ChatGPT-4o, providing to the model a more comprehensive understanding of the nuances involved in identifying and explaining these specific linguistic patterns. The fine-tuned model • • • • • • • aimed to enhance its accuracy and ability to provide clearer, context-sensitive explanations for such complex tasks.
The initial phase of our work focused on generating synthetic phrases for two distinct tasks: identifying age-specific language and detecting sexual content in text.
For the age detection task, we defined the following parameters to categorize the content: For the sexual content detection task, the following parameters were established: Age group: This category helps to classify text based on the perceived age of the speaker. o 13 years or under o 14 -17 years o 18 years or over Sentence complexity: This parameter captures the complexity of the language used, which can vary across age groups.
o very simple o simple o intermediate Use of slang: Identifying whether slang is present helps in determining the likelihood of the text belonging to a younger or older audience.
o Yes o No Sexual allusion: This identifies whether the text contains indirect or direct references to sexual matters.
o Yes o No Presence of emojis: Emojis can often indicate a more casual or suggestive tone, influencing the interpretation of the text.
o Yes o No Reference to digital content: This category captures whether the text mentions digital content, such as images or videos, which may relate to explicit or suggestive material. o None o Image o Video Harmful behavior: This parameter focuses on detecting negative attitudes such as harassment, manipulation, coercion, or objectification within the text.
o Harassment o Manipulation o Coercion o Objectification o None • • • • • •
Explicit language: This measures the level of explicit language used in the text, from “none” to “high”.
o No o Low o Medium o High The next step involved two datasets, one for the age detection task and one for the sexual content detection task, creating a unified collection of synthetic phrases. By generating a total of 500 samples across both tasks, we ensured a diverse representation of content. These samples were carefully annotated based on the parameters outlined above, which enabled us to finetune the model to better identify and explain age-specific language and sexual references in text. Subsequently, an accurate explanation for each phrase in the dataset was generated (using ChatGPT-4o).
Moreover, the explanation of each phrase, along with the previously identified categories, were grouped into a single paragraph. The final dataset, created in JSON format, was divided into a training set (450 examples) and a test set (50 examples) containing:
Instruction: LLM commands o "Analyse the following text and indicate whether the language used belongs to a minor or an adult, explaining:" o "Analyse the following text and indicates whether the language used contains sexual references or porn, explaining:" Input: dataset phrases
Output: explanation for each phrase An example for each prompt is reported below:
Age task o "Instruction": "Analyses the following text and indicates whether the language used belongs to a minor or an adult, explaining:" o "Input": " My dad took away my PlayStation because I took a punishment." o "Output": "This sentence reflects the language of children aged 13 years or younger, as the parental reaction and focus on the PlayStation is typical of this age group. In particular, it is a sentence characterised by an intermediate complexity that does not make use of slang." Sexual task o "Instruction": " Analyses the following text and indicates whether the language used contains sexual references or porn, explaining:" o "Input": "I imagine you in your room right now, make a video for me." o "Output": "The text presents references to digital content, in particular videos.
The imagining of the person in their room, accompanied by the request for a video, evokes a private and sensual image, typical of an erotic context. It is possible to define the attitude in question as harassment. In addition, one can assess the language used as highly suggestive of sexual themes.” Finally, fine-tuning was performed on LLaMAntino (version 3-ANITA-8B-Inst-DPO-ITA) using the previously created dataset.
For the training, we chose LLaMAntino as the pre-trained model. LLaMAntino [
Fine-tuning of the model was performed on Kaggle using the following parameters: • • • • • num train epochs: 10 learning rate: 2e-4, logging steps: 50, optimizer: adamw 8bit, weight decay: 0.01 We obtained the following results (1 epoch corresponds to 56 steps):
Model evaluation was simultaneously performed for both age and sexual tasks. Evaluation of the age task consisted of verifying how accurately the model recognized the types “Age Group”, “Sentence complexity” and “Use of Slang” by comparing them with our ground truth.
Age Group Sentence complexity Use of Slang
Sexual allusion Presence of emojis Reference to content Harmful behavior Explicit language Total digital 25 25 25 75 25 25 25 25 25 125
96,00% 72,00% 88,00% 85,33% Accuracy 100,00% 100,00% 96,00% 68,00% 84,00% 89,60%
24 18 22 64 25 25 24 17 21 112 Evaluation of the sexual task consisted of verifying how accurately the model recognized the types “Sexual allusion”, “Presence of emojis”, “Reference to digital content”, “Harmful behavior” and “Explicit language” by comparing them with our ground truth.
The accuracy metrics were assessed using a test set containing 50 sentences. We decided to give 1 point if the category type matched our ground truth and 0 points if there was no match. Then we aggregated all data and we measured partial and total accuracy.
To show the differences between the fine-tuned model and the baseline (LLaMAntino version 3-ANITA-8B-Inst-DPO-ITA), we executed the two following prompts (respectively for age and sexual task), using two sentences from the test set as an example: •
Analyse the following text and indicate whether the language used belongs to a minor or an adult, explaining: “When will I be able to go out alone without them annoying me?”
Baseline
I believe that the language used in this text belongs to a minor (probably a child around 4-6 years of age) in the care of a parent or an adult, as the text expresses a concern related to personal protection, typical of an age when dependence on parents is still great and awareness of one's own safety is in formation. displaying a more autonomous tone. In particular, it is a The sentence belongs to 14-17 year olds, who seek autonomy and want to show that they are capable without help by sentence characterised by a simple complexity that does not make use of slang. •
Analyse the following text and indicates whether the language used contains sexual references or porn, explaining: “I'd like to see you without those clothes on... !"#$%&'”
Baseline LLaMAntino fine-tuned
I do not consider that the language used in the text contains sexual or pornographic references.
The text expresses a friendly or romantic expression, with a light and non-offensive tone, as the sentence is addressed to someone known and not to an anonymous subject. The presence of an emoticon such as
suggests a playful and non-sexually explicit tone. In general, the text is more akin to a polite expression and not erotic or offensive content.
The phrase emphasises the desire to see the person without clothes, implying an explicit sexual intention. The tone is made more provocative by the emoji !"#$%&'. It is possible to define the attitude in question as objectification. In addition, one can assess the language used as highly racy on issues of a sexual nature.
To test our model in a typical use case, we simulated a chat between two people, one adult and the other a minor, and applied prompts for age recognition and sexual language detection in sequence.
•
Analyse the following text and indicate whether the language used belongs to a minor or an adult, explaining: “User 1: You look too sexy in that photo, but I want to see you without anything. User 2: I don't know... mum says you don't do these things”
The User 1 sentence reflects sexually explicit language typical of an adult context, while the User 2 reflection introduces an ethical concern characteristic of 14-17 year old adolescents. In particular, it is a sentence characterised by an intermediate complexity that does not make use of slang. •
Analyse the following text and indicates whether the language used contains sexual references or porn, explaining: “User 1: You look too sexy in that photo, but I want to see you without anything. User 2: I don't know... mum says you don't do these things”
LLaMAntino fine-tuned
The text contains references to digital content, in particular photos. The request for nudity emphasises a specific interest in explicitly sexual content. The language used is directly related to a context of an erotic nature. It is possible to define the attitude in question as coercion. In addition, one can assess the language used as highly racy on topics of a sexual nature.
As we can notice from the previously provided examples, it is evident that the baseline model generates responses that are more generic, imprecise, and less aligned with our specific goals. In contrast, the fine-tuned model produces more technical and detailed responses, indicating that the fine-tuning process has enhanced its ability to perform the proposed task effectively. Looking ahead, there is still room for improvement. The absence of real-world data has limited the model overall accuracy, as it struggles to capture the nuances of diverse age groups and variations in sexual language. Collaboration with field experts, such as psychologists and sociologists, could help refine the selection of dataset categories and provide more precise advice and explanations tailored to the task.
Additionally, experimenting with more advanced models, such as LLaMAntino with 13B parameters, could further enhance performance. However, it is important to consider the tradeoff between effectiveness and efficiency, as overly complex models may improve accuracy at the expense of practical usability.
This work was partially supported by the following projects: SERICS - “Security and Rights In the CyberSpace - SERICS” (PE00000014) under the MUR National Recovery and Resilience Plan funded by the European Union – NextGenerationEU.
The authors have not employed any Generative AI tools.