by Zaevlad
Audit Findings Dataset is a collection of 23,625 smart-contract security audit findings with titles, descriptions, proof-of-concept code, recommendations and severity. The card calls it raw, semi-structured data that needs cleaning before model training.
DatasetFree
Step 1: Read the card's notes on limitations before using the data.
Step 2: Load the Parquet file with datasets, pandas or polars.
Step 3: Filter bug_poc by length or content to remove placeholder values such as 'no poc'.
Step 4: Deduplicate rows and review the severity labels, using bug_sev_raw if you want your own mapping.
Step 5: Use the cleaned data for defensive research on vulnerability detection or report generation.
No. The card says it is raw, semi-structured data that needs deduplication, label normalization and filtering of low-quality entries first.
About 78.1% of rows (18,443 of 23,625) have a placeholder in bug_poc rather than real code.
The card says to treat it as 'other' until provenance is confirmed, since the underlying audit reports come from third-party researchers.
It is mainly a write-up thoroughness score gated by a coarse severity tier, not a precise importance ranking.
VerifyAPI MCPVerifyAPI checks a single factual claim against the live web and returns a verdict with a verbatim quote from the source. It is reachable over REST and as an MCP server for AI agents.Chatbot Arena ConversationsChatbot Arena Conversations is LMSYS's dataset of 33,000 cleaned conversations with pairwise human preference votes. It was collected on Chatbot Arena in 2023.
GDPvalGDPval is OpenAI's dataset of 220 real-world knowledge tasks across 44 occupations. It supports evaluating AI models on economically valuable work.