The dataset behind DeveloperSkillHubs
Grounded developer skills, mapped from real code.

Explore the Website · Dataset Configurations · Quick Start

CodeSkillBank

CodeSkillBank is a large-scale collection of reusable programming skills grounded in real source-code implementations. It is constructed with Code2Skill, an automated pipeline that transforms implementation evidence into structured procedural knowledge, and powers the accompanying DeveloperSkillHubs website.

The three names refer to different layers:

• DeveloperSkillHubs — the public project website and research presentation;
• CodeSkillBank — the dataset released here;
• Code2Skill — the method used to construct the dataset.

Unlike skills derived only from documentation or agent trajectories, CodeSkillBank records are grounded in concrete source implementations. Each card captures an applicable procedure through its workflow, inputs, outputs, invariants, failure cases, and anti-goals.

At a Glance

The full evidence archive and this Hugging Face release are different resource layers. The release provides a purpose-indexed, intent-deduplicated interface for retrieval, analysis, and downstream agent use.

Why It Matters

Across nine model settings and eight benchmarks, retrieving CodeSkillBank records increases the reported macro-average score from 42.90 to 47.90—an 11.7% relative improvement. Improvements are observed in 57 of 72 protocol-matched model–benchmark evaluations, including all nine reported SWE-bench Verified comparisons.

These results follow the evaluation protocols described in the paper and do not guarantee improvements for every model, task, or individual skill card.

Dataset Configurations

The primary key of cards is purpose_id. Source-level membership can be reconstructed with:

edges.purpose_id = cards.purpose_id

Annotations can be joined to cards using the same purpose_id.

What Is in a Skill Card?

Each row in cards represents one reusable programming purpose after deterministic intent deduplication.

• Intent: task_family, intent_action, intent_target
• Procedure: transfer_title, transfer_summary, workflow
• Interface: inputs, outputs
• Guardrails: invariants, error_cases, anti_goals
• Evidence: support_count, repo_support, representative

Skills with the same normalized purpose—task_family + intent_action + intent_target—are grouped into one card. This is deterministic intent clustering, not embedding-based or pairwise-LLM semantic deduplication.

Construction

Code2Skill builds the bank in four stages:

1. Source-unit selection: identify functions, methods, command entry points, and components containing reusable procedural evidence.
2. Typed extraction: represent source knowledge as atomic, composite, or recurring-pattern skills.
3. Grounding challenge: reconstruct code without exposing the original source body, then compare the reconstruction with the source implementation.
4. Indexing: retain accepted records with provenance and organize them into purpose-level retrieval views.

The reconstruction stage is an LLM-based consistency challenge, not a formal proof of program equivalence.

Pattern Skill Annotations

A confirmed Pattern Skill requires evidence from at least two composite instances in different repositories, a stable sequence of phases, parameterizable variation, and explicit applicability conditions and boundaries.

The release retains supporting evidence identifiers, confidence values, decision reasons, model information, prompt versions, and serialized review traces for auditing.

Loading the Dataset

from datasets import load_dataset

DATASET_ID = "ant-intl/CodeSkillBank"

cards = load_dataset(DATASET_ID, "cards", split="train")
edges = load_dataset(DATASET_ID, "edges", split="train")
patterns = load_dataset(DATASET_ID, "pattern_annotations", split="train")

Select confirmed recurring patterns with:

confirmed_patterns = patterns.filter(lambda row: row["is_pattern"])

Intended Uses

CodeSkillBank supports:

• procedural-memory retrieval for coding agents;
• software implementation, debugging, and code review;
• retrieval-augmented code generation;
• source-grounded skill synthesis and analysis;
• reusable workflow and implementation-pattern discovery;
• training and evaluation of software-engineering agents.

Records can be used as model context, retrieval documents, planning references, or auditable intermediate representations. The dataset is not a collection of executable source files.

Limitations

CodeSkillBank records are textual abstractions of source implementations. They may contain omissions, model-generated inaccuracies, outdated assumptions, or overly broad applicability conditions. Retrieved procedures should be validated before use in security-sensitive, safety-critical, or production environments.

This data-only release does not distribute the Code2Skill construction pipeline or original source bodies. Internal absolute source paths have been removed.

Licensing and Provenance

The current source metadata does not yet include fully verified upstream URLs, commit hashes, and repository-level license information for every record. The dataset card therefore uses license: other until the provenance audit and dataset-level licensing decision are complete.

Additional schema, statistics, and integrity information is available in:

• metadata/data_dictionary.md
• metadata/schema.json
• metadata/statistics.json
• metadata/checksums.sha256