Language / 语言: 中文 | English

This directory brings together the PCART-LLM code changes, evaluation dataset, ground-truth data, experiment outputs, and evaluation scripts.

PCART-LLM/
├── PCART-LLM-Code/
├── PCART-LLM-Evaluation/
├── PCBench-LLM-Distilled/
├── PCBench-LLM-Distilled-groundTruth/
└── PCBench-LLM-Distilled-Result/

• PCART-LLM-Code: stores the side-by-side snapshots of the original PCART implementation and the locally modified implementation. It contains PCART-original/, PCART-modified/, and patches/, showing the LLM-based findDiffer changes, batch experiment support, and patches relative to the original implementation.
• PCBench-LLM-Distilled: stores the API-level test set for evaluating the capabilities of parameter-change methods. The current Benchmark/ directory contains 284 retained samples from 29 libraries, together with the sample manifest and scripts for environment download and JSON configuration generation.
• PCBench-LLM-Distilled-groundTruth: stores the DictOnly ground truth for PCBench-LLM-Distilled. It organizes 284 expected-result JSON files by the library / api@version_range / sample hierarchy and is used for strict ordered comparison against experiment outputs.
• PCBench-LLM-Distilled-Result: stores result files produced by different models or baselines on the PCBench-LLM-Distilled dataset. It includes result groups such as claude/, deepseek/, gemini/, minimax/, qwen/, and default/, together with the corresponding semantic split reports.
• PCART-LLM-Evaluation: stores standalone evaluation helper scripts, including the result JSON cleanup script and the semantic-related / semantic-unrelated split comparison report generator.

PCART-LLM is licensed under the GNU Affero General Public License v3.0. See LICENSE for details.