kumarde / llm-content-mod

Datasets and code for studying LLMs and Content Moderation

llm-content-mod

Datasets and code for studying LLMs and Content Moderation

Datasets

• The data/ folder contains all the subreddit specific data and toxic content data we curated to conduct our experiments
• rule_moderation/ contains balanced datasets for each subreddit we studied. In particular:
  • subreddit_rules_w_description.jsonl contains a description and the rule-string we used for prompting each LLM
  • subreddit_balanced_datasets contains the actual balanced data files for each subreddit
• toxic_content/ contains 10k_balanced_sample.jsonl, which is the data file with balanced toxic/nontoxic comments curated for our toxic content evaluation

Scripts

• The scripts/ folder has all the base scripts we used to run our data through various commercial models.
• scripts/is_toxic_*.py has all the scripts for toxic content detection
• scripts/rule_based_moderation.py has the code for evaluating rule-based decisions per subreddit