Datasets:

cezar-sas
/

code-theseus

license: cc-by-4.0
pretty_name: code-theseus
task_categories:
  - text-generation
language:
  - en
size_categories:
  - 10K<n<100K
tags:
  - software-engineering
  - requirements-engineering
  - software-architecture
  - reverse-engineering
  - c4-model
  - code-generation
  - coding-agents
  - llm
  - synthetic
configs:
  - config_name: documents
    data_files: documents/train.parquet
    default: true
  - config_name: repos
    data_files: repos/train.parquet

🚢 code-theseus

The design behind 10,000+ open-source repositories: intent, requirements and architecture, aligned with the code, ready for training coding agents.

📦 10,813	📄 21,623	🌐 4	🤖 2
repositories	document sets	languages: Java, Python, C/C++, C#	generator LLMs

🧭 Why this dataset

Coding agents increasingly design before they code, but almost no data connects code to the requirements and architecture behind it: real projects rarely keep these documents, and when they do they are outdated or private. code-theseus recovers them from the source code, at scale.

📦 What's inside

Every repository comes with three aligned documents, generated from its dependency graph and key project files:

	Document	What it captures
🎯	`intent`	The "initial prompt": mission, problem, capabilities, user workflows, quality expectations
📋	`requirements`	`REQ-FUNC-` / `REQ-NFR-` items with statement, rationale, acceptance criteria, verification method and priority
🏗️	`architecture`	A C4 design (context, container, component) with Mermaid C4 diagrams

…plus the package-level dependency graph each document was generated from, and a link to the exact commit it describes. No source code is redistributed: check out url at commit_sha to recover it.

🔁 Does it carry the design?

In the paper's Ship of Theseus test, a coding agent rebuilt Commit0 libraries from these documents alone. With requirements and architecture, it built a module structure significantly closer to the original than with intent alone (module F1 0.54 → 0.63), at the same test pass rate.

🏢 Use it for

Fine-tuning models that write specs from code (graph → intent, requirements, architecture) or code from specs.
Spec-driven coding agents. Realistic requirements and architecture briefs as agent inputs.
Evaluating the intermediate steps of multi-agent development, not just the final code.
Research on requirements engineering, software architecture recovery and traceability.

🚀 Quick start

from datasets import load_dataset

docs = load_dataset("anonymous/code-theseus", "documents", split="train")
repos = load_dataset("anonymous/code-theseus", "repos", split="train")
graphs = dict(zip(repos["repo_id"], repos["arcan_graph"]))
graph = graphs[docs[0]["repo_id"]]  # input graph for the first document

🗂️ Configs

`documents` (21,623 rows): one row per (repository, generator)

column	description
`repo_id`	`owner\|repo`; join key to `repos`
`generator`	`gemma-4-31b-it` / `qwen3.5-122b-a10b`
`checkpoint`	Hugging Face checkpoint served with vLLM
`intent`, `requirements`, `architecture`	generated Markdown; null if missing, truncated or empty (56 docs)
`reasoning_*`	the generator's thinking trace (Qwen only, null otherwise)

`repos` (10,813 rows): one row per repository

column	description
`repo_id`	`owner\|repo`
`branch`	`main` (GitHub corpus) or `commit0` (Commit0 libraries)
`url`	GitHub URL
`commit_sha`	analysed commit; documents describe the repository at this commit
`language`, `description`, `stars`	GitHub metadata at collection time
`arcan_graph`	package-level dependency graph from Arcan: `nodes` (`vertex_id`, `name`, `component_type`, `fan_in`, `fan_out`, `lines_of_code`, `pagerank`) and weighted `edges` (`source`/`target` are `vertex_id`s)

📜 License

CC BY 4.0. This covers the generated documents and dependency graphs. The original repositories keep their own licenses.