Prompt Injection Analysis Dataset with 48,090 Multi-Source Chat Examples

48,090 chat conversations for training large language models to analyze prompt injection techniques, jailbreak patterns, and security defenses. The dataset was created by SkywardNomad92 and last updated on February 12, 2026. It aggregates examples from seven sources including mosscap, open_prompt_injection, and jailbreakhub.

Use Cases

• Fine-tuning LLMs for prompt injection detection based on the described chat conversation format.
• Analyzing jailbreak attack patterns based on the aggregated examples from multiple security sources.
• Training classifiers for LLM security defenses based on the dataset's focus on adversarial techniques.

Strengths

• Aggregates 48,090 examples from seven distinct sources, providing a multi-perspective collection.
• Each example is structured as a 3-message chat conversation, offering a consistent format for analysis.
• The largest single source, mosscap, contributes 20,000 examples (41.6% of the total).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but specific file formats, sample data, and column details are unavailable.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Hugging Face user SkywardNomad92

Collection Method

Aggregated from seven public sources: mosscap, open_prompt_injection, safeguard, jailbreakhub, jailbreak_classification, deepset, and chatgpt_jailbreaks.

Time Range

null

Freshness

Last updated 2026-02-12 20:52:32; freshness should be verified.

Geography

null