Self-Prompt: Optimized System Prompts for Code Generation LLMs

Self-Prompt project data contains automatically discovered optimal system prompts for code generation large language models. The dataset includes base prompts, optimized soft-prompt tokens, and final assembled prompts for models evaluated on HumanEval, MBPP, and Eval_PLUS benchmarks. It was authored by liangdongzheng zhu and last updated on 2026-06-02.

Use Cases

• Benchmarking code generation models based on the provided HumanEval, MBPP, and Eval_PLUS prompts
• Applying gradient-guided discrete prompt optimization techniques based on the described methodology
• Analyzing the structure of assembled system prompts (base + extra tokens) for code generation tasks
• Reproducing or extending soft-prompt optimization experiments for LLMs

Strengths

• Dataset is associated with three established code generation benchmarks: HumanEval, MBPP, and Eval_PLUS
• Includes multiple prompt components: base prompts, optimized soft-prompt tokens, and final assembled prompts
• Released under a permissive CC-BY-4.0 license for open use and modification

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download
• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment

Provenance

Source

figshare

Collection Method

Automatically discovered via soft-prompt optimization and evaluation on code generation benchmarks.

Freshness

Last updated 2026-06-02 05:40:08; freshness should be verified