SWHForest (v0.2.0)

SWHForest lets you filter, search by AST structure, and detect clones across 52.1M files (340.6M functions/classes in the 10 largest languages of HuggingFaceTB/stack-edu, revision eeec5ca) before downloading source code from the Software Heritage Archive (SWH).

No source text is included. Every row is keyed by blob_id (sha1_git); surviving candidate files are fetched on demand from https://softwareheritage.s3.amazonaws.com/content/<blob_id>.

• Code, CLI & Python package: https://github.com/halykoss/swhforest
• Column-level schema: SCHEMA.md

---

Download & Install

Clone the code repository and pull either the lightweight 100k Python/JS benchmark sample (catalog/, ~88 MB) or any language release (data/release/<lang>/):

git clone https://github.com/halykoss/swhforest.git && cd swhforest
pip install -r requirements.txt && pip install -e ".[all-languages,dev]"

# Download the 100k sample or one language release into the repository
hf download halykoss/swhforest --repo-type dataset --include "catalog/*" --local-dir .
hf download halykoss/swhforest --repo-type dataset --include "data/release/rust/*" --local-dir .

---

Quick Examples

1. Search by AST Pattern (swhforest.search + ast-grep)

Single-relation patterns (exact=True) are answered directly by the local index with 0 downloads; multi-part patterns prune the corpus locally and verify surviving candidates with ast-grep.

import swhforest
from swhforest import Pattern

# A) Single-block query -> exact answer from the local index (0 files downloaded)
p_exact = Pattern(type="function_item", contains=Pattern(type="unsafe_block"))
res = swhforest.search("data/release/rust", p_exact, language="rust")
print(res["exact"], res["n_candidates"])  # True, 29527

# B) Multi-block query -> local index pruning + live SWH fetch + ast-grep verification
call_clone = Pattern(type="call_expression", fields={
    "function": Pattern(type="field_expression", fields={
        "field": Pattern(type="field_identifier", text="clone")
    })
})
p_loop = Pattern(type="for_expression", contains=call_clone)
res = swhforest.search("data/release/rust", p_loop, language="rust", verify_limit=10)
print(res["n_candidates"], len(res["verified_blob_ids"]))

---

2. Find Renamed Clones & Screen Benchmarks (swhforest.find_clones)

Every function, method, and class has a 64-bit rename-invariant AST Merkle hash (shape). Pass any snippet (e.g., a HumanEval/MBPP solution) to find structurally identical definitions across the corpus and slice their bytes directly from SWH:

import swhforest

code = """
def has_close_elements(numbers, threshold):
    for idx, elem in enumerate(numbers):
        for idx2, elem2 in enumerate(numbers):
            if idx != idx2:
                if abs(elem - elem2) < threshold:
                    return True
    return False
"""

matches = swhforest.find_clones("data/release/python", code, language="python", min_nodes=20, limit=5)
for m in matches:
    print(f"{m['repo']}:{m['path']} -> blob {m['blob_id']} [{m['start_byte']}:{m['end_byte']}]")
    print(swhforest.slice_definition(m["blob_id"], m["start_byte"], m["end_byte"]))

---

3. Curate Subsets & Export a Reproducible View (swhforest.curate_view)

Filter files by their 17 syntactic metrics in catalog.parquet (file size, docstrings, function/class counts, AST depth, parse errors) and export a JSON manifest (blob_id list + SHA-256 checksum):

import swhforest

view = swhforest.curate_view(
    "data/release/python",
    where="bytes > 10000 AND has_docstring = true AND (n_functions + n_methods) >= 5",
    out_manifest="curated_python.json",
)
print(f"{view['n_blobs']:,} files ({view['total_bytes'] / 1e9:.2f} GB) | SHA-256: {view['blob_ids_sha256'][:16]}...")

assert swhforest.verify_view("data/release/rust", "data/release/rust/view_release.json")

Remote Parquet Queries: You can also query the Parquet tables directly on Hugging Face without downloading them first:

import duckdb
duckdb.sql("""
    SELECT count(*), sum(bytes)
    FROM 'hf://datasets/halykoss/swhforest/data/release/python/catalog.parquet'
    WHERE bytes > 10000 AND has_docstring AND (n_functions + n_methods) >= 5
""").show()

---

Dataset Layout & Summary (v0.2.0)

Deterministic 40% SHA-1 prefix sample (blob_id < 66666666) of deduplicated 200 B–200 kB files across the 10 largest languages in Stack-Edu:

data/release/<language>/       python, javascript, java, cpp, csharp, php, c, typescript, go, rust
  rows.parquet                 1 row per unique blob_id: row id, blob_id, representative repo/path/license/edu_score
  catalog.parquet              1 row per file: 17 syntactic metrics (AST size/depth, docstrings, definitions, imports)
  functions.parquet            1 row per function/method/class: byte span [start_byte, end_byte], nodes, 64-bit shape hash
  shapes.parquet               1 row per (kind, shape) cluster: count, mean_nodes, sample blob_id[start:end]
  structure-<k>.parquet        inverted AST feature index shards (2M files each)
  view_release.json            reproducible view manifest (predicate, pinned parser versions, SHA-256 checksum)
  selection.json, ingest_metrics.json, missing.txt
catalog/                       100k Python/JavaScript benchmark sample (used by examples 01-03)

Language	Files (blob_id)	Definitions	Source Text (Not Stored)	Tables (GB)	Index (GB)
Java	17,429,408	135,578,481	53.94 GB	6.15	7.52
Python	9,481,411	43,406,601	27.10 GB	2.82	3.16
C++	6,453,978	39,997,408	17.31 GB	2.25	2.71
JavaScript	5,173,711	39,520,638	13.77 GB	2.04	1.77
C#	4,560,555	30,471,416	13.74 GB	1.65	1.81
PHP	3,904,782	20,813,360	11.24 GB	1.17	0.86
C	2,266,959	8,370,541	6.21 GB	0.64	0.74
TypeScript	1,673,249	12,517,615	3.85 GB	0.66	0.66
Go	752,530	4,823,599	1.84 GB	0.26	0.26
Rust	448,639	5,149,143	2.12 GB	0.23	0.29
Total	52,145,222	340,648,802	151.12 GB	17.87	19.77

---

Authors, License & Citation

• Authors: Federico Pennino, Andrea Gurioli, Maurizio Gabbrielli (University of Bologna), and Paolo Ferragina (Sant'Anna School of Advanced Studies, Pisa).
• Dataset tables, index & manifests: CC-BY-4.0 · Code: MIT (halykoss/swhforest) · Source blobs: fetched on demand from Software Heritage under their original upstream licenses.
• Opt-out and takedowns: SWHForest follows upstream removals from The Stack v2 / Stack-Edu and Software Heritage; withdrawn blob_ids are dropped in subsequent releases.

@misc{pennino2026swhforest,
  author        = {Pennino, Federico and Gurioli, Andrea and Gabbrielli, Maurizio and Ferragina, Paolo},
  title         = {{SWHForest}: A Structural Curation and Search Layer for Large-Scale Code Corpora},
  year          = {2026},
  eprint        = {2611.11439},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SE},
  url           = {https://arxiv.org/abs/2611.11439}
}