---
title: Organization Card
emoji: 🚀
colorFrom: blue
colorTo: green
sdk: static
pinned: false
license: mit
---

# Precise Debugging Benchmarking (PDB)

💻 [Code](https://github.com/anon-pdb-2026/anon-pdb-code)  · 
🌐 [Project page](https://anon-pdb-2026.github.io)  · 
🏆 [Leaderboard](https://anon-pdb-2026.github.io/leaderboard.html)

> *Anonymous release for NeurIPS 2026 Datasets & Benchmarks review.*

**PDB** is an automatic pipeline that turns any coding dataset into a
*debugging* benchmark with fine-grained metrics. Beyond binary unit-test
scores, PDB evaluates a debugger with **edit-level precision** (did the model
touch only the lines it had to?) and **bug-level recall** (did it fix every
fault?). This rewards targeted fixes and penalizes the regeneration behavior
frontier LLMs often fall back on.

## Released datasets

| Dataset | Size | Bug granularity | Notes |
|---|---|---|---|
| [PDB-Single](https://huggingface.co/datasets/anon-pdb/PDB-Single) | 5,751 | single line | hard subset of single-line bug examples |
| [PDB-Wild](https://huggingface.co/datasets/anon-pdb/PDB-Wild) | 484 | line blocks | multi-line synthesized bugs, real-world repository bugs |

All datasets ship a unified schema with `source_dataset` provenance, ground-truth
edit scripts (`gt_diff`), and Orthogonal Defect Classification labels.