datatrove-rs 0.1.0__tar.gz

datatrove_rs-0.1.0/Cargo.lock

@@ -0,0 +1,320 @@
+# This file is automatically @generated by Cargo.
+# It is not intended for manual editing.
+version = 4
+
+[[package]]
+name = "aho-corasick"
+version = "1.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c982642fa9e8606056828ee9a8505737230110bb1099153c79efe865c59d12ba"
+dependencies = [
+ "memchr",
+]
+
+[[package]]
+name = "bit-set"
+version = "0.8.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "08807e080ed7f9d5433fa9b275196cfc35414f66a0c79d864dc51a0d825231a3"
+dependencies = [
+ "bit-vec",
+]
+
+[[package]]
+name = "bit-vec"
+version = "0.8.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5e764a1d40d510daf35e07be9eb06e75770908c27d411ee6c92109c9840eaaf7"
+
+[[package]]
+name = "datatrove-core"
+version = "0.1.0"
+dependencies = [
+ "hashbrown",
+ "regex",
+ "rustc-hash",
+ "serde",
+ "serde_json",
+ "sha1_smol",
+ "spacy-tokenizer-rs",
+ "xxhash-rust",
+]
+
+[[package]]
+name = "datatrove-py"
+version = "0.1.0"
+dependencies = [
+ "datatrove-core",
+ "pyo3",
+]
+
+[[package]]
+name = "fancy-regex"
+version = "0.18.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e1e1dacd0d2082dfcf1351c4bdd566bbe89a2b263235a2b50058f1e130a47277"
+dependencies = [
+ "bit-set",
+ "regex-automata",
+ "regex-syntax",
+]
+
+[[package]]
+name = "hashbrown"
+version = "0.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ed5909b6e89a2db4456e54cd5f673791d7eca6732202bbf2a9cc504fe2f9b84a"
+
+[[package]]
+name = "heck"
+version = "0.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea"
+
+[[package]]
+name = "itoa"
+version = "1.0.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
+
+[[package]]
+name = "libc"
+version = "0.2.189"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3eaf3ede3fee6db1a4c2ee091bf8a8b4dccdc6d17f656fb07896ee72867612f2"
+
+[[package]]
+name = "memchr"
+version = "2.8.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "cf8baf1c55e62ffcace7a9f06f4bd9cd3f0c4beb022d3b367256b91b87513d98"
+
+[[package]]
+name = "once_cell"
+version = "1.21.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50"
+
+[[package]]
+name = "portable-atomic"
+version = "1.15.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "05c8b63e8d9609db387f0324918f81d68fe27748f084ef092fb35954d0539a85"
+
+[[package]]
+name = "proc-macro2"
+version = "1.0.107"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "985e7ec9bb745e6ce6535b544d84d6cd6f7ad8bd711c398938ae983b91a766d9"
+dependencies = [
+ "unicode-ident",
+]
+
+[[package]]
+name = "pyo3"
+version = "0.29.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4688ddedf473e32662b9b067670129a8afb8c18e351482c70d62ba4a88171e8b"
+dependencies = [
+ "libc",
+ "once_cell",
+ "portable-atomic",
+ "pyo3-build-config",
+ "pyo3-ffi",
+ "pyo3-macros",
+]
+
+[[package]]
+name = "pyo3-build-config"
+version = "0.29.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f41027e41b4bd03f6e60f9f417fe24a6341a6bb744edd62b6f709f2a52ea30e9"
+dependencies = [
+ "target-lexicon",
+]
+
+[[package]]
+name = "pyo3-ffi"
+version = "0.29.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e591a95526fead067432c3b3a33fc74770b87b1e04e73671090d9c2055a2b327"
+dependencies = [
+ "libc",
+ "pyo3-build-config",
+]
+
+[[package]]
+name = "pyo3-macros"
+version = "0.29.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "73225868fc1cd84eef2c3c230ddb91273bf1de46aeb8a4248da76d32a0924a1c"
+dependencies = [
+ "proc-macro2",
+ "pyo3-macros-backend",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "pyo3-macros-backend"
+version = "0.29.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "571575aa3749fa6216757dd47d2a3e7ef360f329a40f0666a9fbd14889024952"
+dependencies = [
+ "heck",
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+]
+
+[[package]]
+name = "quote"
+version = "1.0.47"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1fbf4db142a473a8d80c26bbf18454ed458bf8d26c8219c331daecfdbd079001"
+dependencies = [
+ "proc-macro2",
+]
+
+[[package]]
+name = "regex"
+version = "1.13.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f020237b6c8eed93db2e2cb53c00c60a8e1bc73da7d073199a1180401450218d"
+dependencies = [
+ "aho-corasick",
+ "memchr",
+ "regex-automata",
+ "regex-syntax",
+]
+
+[[package]]
+name = "regex-automata"
+version = "0.4.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ad8553b9b26413251cbf30e620595c7a41b3887f03da04579c0e6b0d6a06b4b2"
+dependencies = [
+ "aho-corasick",
+ "memchr",
+ "regex-syntax",
+]
+
+[[package]]
+name = "regex-syntax"
+version = "0.8.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d6f6ff9a378485b298a5286656da665ba74413d36db0979633275d2e708145d4"
+
+[[package]]
+name = "rustc-hash"
+version = "2.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6b1e7f9a428571be2dc5bc0505c13fb6bf936822b894ec87abf8a08a4e51742d"
+
+[[package]]
+name = "serde"
+version = "1.0.229"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4148590afebada386688f18773da617792bf2ef03ffc1e4cbd2b1d45b023e0ba"
+dependencies = [
+ "serde_core",
+ "serde_derive",
+]
+
+[[package]]
+name = "serde_core"
+version = "1.0.229"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "67dca2c9c51e58a4791a4b1ed58308b39c64224d349a935ab5039aa360942a48"
+dependencies = [
+ "serde_derive",
+]
+
+[[package]]
+name = "serde_derive"
+version = "1.0.229"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e7a5d71263a5a7d47b41f6b3f06ba276f10cc18b0931f1799f710578e2309348"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 3.0.6",
+]
+
+[[package]]
+name = "serde_json"
+version = "1.0.151"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c841b55ecdae098c80dcae9cf767f6f8a0c2cdb3416bbef72181df4d0fe73f14"
+dependencies = [
+ "itoa",
+ "memchr",
+ "serde",
+ "serde_core",
+ "zmij",
+]
+
+[[package]]
+name = "sha1_smol"
+version = "1.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bbfa15b3dddfee50a0fff136974b3e1bde555604ba463834a7eb7deb6417705d"
+
+[[package]]
+name = "spacy-tokenizer-rs"
+version = "0.1.0"
+dependencies = [
+ "fancy-regex",
+ "hashbrown",
+ "regex",
+ "regex-syntax",
+ "rustc-hash",
+ "serde",
+ "serde_json",
+]
+
+[[package]]
+name = "syn"
+version = "2.0.119"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "872831b642d1a07999a962a351ed35b955ea2cfc8f3862091e2a240a84f17297"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "unicode-ident",
+]
+
+[[package]]
+name = "syn"
+version = "3.0.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8593e8e72159ed2257d083c7a454a85cbf854f37a0966d8d483aff8c8a3ebcee"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "unicode-ident",
+]
+
+[[package]]
+name = "target-lexicon"
+version = "0.13.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "adb6935a6f5c20170eeceb1a3835a49e12e19d792f6dd344ccc76a985ca5a6ca"
+
+[[package]]
+name = "unicode-ident"
+version = "1.0.26"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d245f478577f809a851594d02313b640fb437e0bb33866753cff937863096954"
+
+[[package]]
+name = "xxhash-rust"
+version = "0.8.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "aee1b19627c7c60102ab80d3a9cbe18de90bfe03bfa6c3715447681f0e8c8af6"
+
+[[package]]
+name = "zmij"
+version = "1.0.23"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "29666d0abbfad1e3dc4dcf6144730dd3a3ab225bbbdac83319345b1b44ccfc1b"

datatrove_rs-0.1.0/Cargo.toml

@@ -0,0 +1,15 @@
+[workspace]
+resolver = "2"
+members = ["crates/spacy-tokenizer", "crates/datatrove-core", "crates/datatrove-py"]
+
+[workspace.dependencies]
+regex = "1.13"
+rustc-hash = "2"
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+
+[profile.release]
+opt-level = 3
+lto = "thin"
+codegen-units = 1
+debug = 1

datatrove_rs-0.1.0/PKG-INFO

@@ -0,0 +1,207 @@
+Metadata-Version: 2.4
+Name: datatrove-rs
+Version: 0.1.0
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Operating System :: MacOS
+Classifier: Operating System :: Microsoft :: Windows
+Classifier: Operating System :: POSIX :: Linux
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Rust
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Text Processing :: Linguistic
+Requires-Dist: datatrove>=0.10.1,<0.11
+Requires-Dist: numpy
+Requires-Dist: regex
+Requires-Dist: pytest ; extra == 'test'
+Requires-Dist: pyarrow ; extra == 'test'
+Requires-Dist: spacy>=3.8,<3.9 ; extra == 'test'
+Requires-Dist: xxhash<4 ; extra == 'test'
+Requires-Dist: tokenizers ; extra == 'test'
+Requires-Dist: trafilatura-rs[trafilatura]>=0.1,<0.2 ; extra == 'trafilatura'
+Provides-Extra: test
+Provides-Extra: trafilatura
+Summary: Drop-in Rust implementations of datatrove's English Gopher, C4, FineWeb filters and MinHash stage 1
+Keywords: datatrove,llm,pretraining,data-filtering,deduplication,minhash,fineweb,rust
+Author: Hexay
+License: Apache-2.0
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown; charset=UTF-8; variant=GFM
+Project-URL: Homepage, https://github.com/Hexay/datatrove-rs
+Project-URL: Issues, https://github.com/Hexay/datatrove-rs/issues
+Project-URL: Repository, https://github.com/Hexay/datatrove-rs
+
+# datatrove-rs
+
+Drop-in Rust implementations of [datatrove](https://github.com/huggingface/datatrove)'s FineWeb filter chain and
+MinHash stage 1, for English. The outputs match datatrove's exactly: the same keep/drop decisions, drop reasons,
+rewritten `doc.text`, stats and signature files.
+
+| datatrove step | Rust-backed replacement |
+|---|---|
+| `pipeline.filters.GopherRepetitionFilter` | `datatrove_rs.GopherRepetitionFilter` |
+| `pipeline.filters.GopherQualityFilter` | `datatrove_rs.GopherQualityFilter` |
+| `pipeline.filters.C4QualityFilter` | `datatrove_rs.C4QualityFilter` |
+| `pipeline.filters.FineWebQualityFilter` | `datatrove_rs.FineWebQualityFilter` |
+| `pipeline.dedup.MinhashDedupSignature` | `datatrove_rs.MinhashDedupSignature` |
+| `pipeline.extractors.Trafilatura` | `datatrove_rs.extractors.Trafilatura`, with the `[trafilatura]` extra |
+
+The Trafilatura step runs [trafilatura-rs](https://github.com/Hexay/trafilatura-rs), a byte-exact Rust port of
+trafilatura 1.11 (`pip install "datatrove-rs[trafilatura]"`). Without it, datatrove's own Trafilatura runs. Configs
+trafilatura-rs doesn't cover fall back to Python per step. On the FineWeb recipe, where Trafilatura is ~85% of CPU,
+this takes the whole stage from 87.7 s to 17.3 s on 5,000 Common Crawl records (5.1x, identical output; see
+`docs/fineweb-profile.md`).
+
+Each class subclasses the datatrove step and takes the same arguments. `filter()` (MinHash: `get_shingles` /
+`get_signature`) is replaced. The filters also override `run()` with an inlined copy of `BaseFilter.run()` whose stat
+bookkeeping is cheaper but bit-for-bit identical (see `python/datatrove_rs/_run.py`); a subclass that overrides
+`filter`, `filter_batch`, `stat_update`, `update_doc_stats` or `track_time` gets datatrove's `run()` instead.
+
+## Adopting
+
+`pip install datatrove-rs`, then pick one; every other datatrove step keeps running datatrove's own code.
+
+1. **No code changes.** Set `DATATROVE_RS=1` (an installed `.pth` hook calls `patch()` at interpreter startup, so
+   executor worker processes get it too), or run the unmodified script with `python -m datatrove_rs pipeline.py`.
+2. **One line.** `import datatrove_rs; datatrove_rs.patch()` before building the pipeline. datatrove's own import
+   paths (`datatrove.pipeline.filters.GopherQualityFilter`, ...) then hand out the Rust classes; `unpatch()` undoes it.
+3. **Explicit imports, no patching.** `datatrove_rs.<path>` mirrors every `datatrove.<path>` module with the Rust
+   classes swapped in, so replacing `datatrove.` with `datatrove_rs.` in the imports is enough:
+
+```python
+from datatrove_rs.pipeline.filters import GopherQualityFilter, URLFilter  # Rust, datatrove's own
+from datatrove_rs.pipeline.readers import WarcReader                       # datatrove's own
+```
+
+The Rust classes keep `__module__ = "datatrove_rs.*"`, so pickled pipelines (Slurm, local multiprocessing) load them
+on workers whether or not the worker patched. Unsupported languages fall back per step, as described below.
+
+Requires datatrove `>=0.10.1,<0.11` and Python 3.10+ (one abi3 wheel). English needs no spaCy install; spaCy is
+only used for the non-English fallback and, when installed, to reproduce the `IS_ALPHA` quirk below.
+
+## Speed
+
+Single thread, 1,000 FineWeb docs (3.0 MB), median of 3 runs (`tools/dev.sh bench <fineweb.parquet> 1000 3`),
+Ryzen 7 9700X, datatrove 0.10.1, Python 3.13:
+
+| Step | datatrove docs/s | datatrove_rs docs/s | Speedup |
+|---|---|---|---|
+| GopherRepetition | 195 | 16,987 | 87x |
+| GopherQuality | 267 | 21,087 | 79x |
+| C4Quality | 268 | 13,119 | 49x |
+| FineWebQuality | 268 | 23,063 | 86x |
+| Chain (`run()`, all 4 filters) | 99 | 6,831 | 69x |
+| MinHash stage 1 | 269 | 12,780 | 48x |
+
+The machine was under heavy load from other processes, so absolute numbers swing between runs; compare the ratio.
+All filters on a thread share one tokenizer state and the last doc's tokens (see `words.rs`).
+
+Where the time goes (`profile <docs.bin> [reps] [stage]`, `tools/icount.sh`): the spaCy-compatible tokenizer still
+dominates (~45 µs/doc streamed). It runs about three times per doc in the chain: whole doc, C4's per-line sentence
+split, and FineWeb on C4's rewritten text (mostly chunk-cache hits). Python `run()` plumbing is about
+60% cheaper than datatrove's `BaseFilter.run()` (`tools/profile_chain.py <fineweb.parquet>`: 99k vs 249k Python calls
+for the 4-step chain over 1,000 docs).
+
+## Parity
+
+Every check compares against datatrove running unmodified in the same environment.
+
+- `tools/dev.sh parity <fineweb.parquet> --real 6000 --fuzz 3000 --random-configs 3` runs 10,500 docs: 6,000 real
+  FineWeb docs, 1,500 perturbed copies and 3,000 fuzzed docs (Unicode edge cases, separators, repetition). It covers
+  9 fixed filter configs, 12 random-threshold configs and 5 MinHash configs (xxhash 32/64, sha1 32/64, 3-gram with
+  4x6 buckets, dates/raw normalisation). It compares:
+  - the per-doc result, or the raised exception type;
+  - the rewritten `doc.text`;
+  - the `run()` output, every stat field (`repr`, so int vs float counts), the time-stat count and unit, and what
+    the exclusion writer receives, at batch sizes 1 to 5;
+  - the MinHash signature files, byte for byte.
+
+  Last run (`--real 4000 --fuzz 2000 --random-configs 3`): **0 mismatches across all 26 cases.** Pass `--workers N --cache DIR` to parallelise and cache the
+  datatrove side.
+- `tools/dev.sh test` (pytest) runs the parity cases on a smaller corpus. It also covers `simplify_text`, word
+  splitting and sentence splitting on 2,500 random Unicode strings across 17 normalisation configs, the fallback for
+  non-English languages, and pickling.
+
+### How exactness is kept, and its limits
+
+- **Unicode tables.** Python's `str.isspace`/`isalpha`/`lower`, `unicodedata` NFD and categories, `re`'s `\d`/`\s`,
+  and the `regex` module's `\p{Nd}` are exported from Python into `crates/datatrove-core/data/py_unicode.json` by
+  `tools/export_pytext.py`. The shipped tables come from Python 3.13 (Unicode 15.1). On import, the package warns if
+  the running Python uses a different Unicode version, since results could then differ on rare characters.
+- **Language.** The Rust path runs only when datatrove would use a plain `spacy.blank("en")` tokenizer. That covers
+  `eng` and every language code datatrove maps to it. Other languages run datatrove's Python code unchanged.
+- **The `IS_ALPHA` token.** Inside `memory_zone`, spaCy 3.8 raises an error on the token `IS_ALPHA` once the same
+  process has seen that token before, and datatrove then retokenizes the text without it. The result depends on
+  process history, so when spaCy is installed any doc containing `IS_ALPHA` runs through datatrove itself.
+  Without spaCy, datatrove could not run at all and Rust handles these docs as spaCy would on first sight.
+- **Reproduced quirks.**
+  - `Counter.most_common` tie-breaking;
+  - the `"".join` n-gram keys in `find_all_duplicate`;
+  - `ZeroDivisionError` on degenerate docs;
+  - the one-sentence result for empty text;
+  - numpy `uint64` wrap-around in the MinHash permutation.
+
+### datatrove 0.10.0 issues found along the way
+
+- `utils/text.py` imports `regex` without declaring it (this package depends on `regex` for that reason).
+- The Python MinHash passes `str` to `xxhash`, which xxhash 4 rejects. The reference side needs `xxhash<4`; this
+  package's MinHash is unaffected.
+- `importlib.metadata` is used without being imported, and `pipeline.dedup` requires `tokenizers`. For the latter,
+  `datatrove_rs.MinhashDedupSignature` is imported lazily, so filter-only installs don't need `tokenizers`.
+
+## Layout
+
+```
+crates/spacy-tokenizer   spaCy English tokenizer port (prefix/suffix/infix rules, specials, URL matcher, chunk cache)
+crates/datatrove-core    filters, MinHash, Python-compatible str/Unicode helpers, `profile` binary
+crates/datatrove-py      PyO3 bindings (datatrove_rs._core)
+python/datatrove_rs      datatrove subclasses and fallback logic
+tools/                   dev.sh, parity/bench harnesses, fuzzers, Unicode table exporter
+tests/                   pytest parity suite
+```
+
+## Development
+
+```
+uv venv -p 3.13 .venv && uv pip install -p .venv/bin/python maturin -e ".[test]"
+tools/dev.sh build                 # maturin develop --release
+tools/dev.sh test                  # pytest
+tools/dev.sh parity <parquet> ...  # large parity run
+tools/dev.sh bench <parquet> [n] [reps]
+tools/icount.sh <docs.bin> <stage>...                  # instructions + cycles/doc per `profile` stage
+crates/spacy-tokenizer/tools/parity.sh <docs.bin>      # tokenizer vs spaCy, token spans
+```
+
+Use Python 3.13 so the running Unicode version matches the shipped tables. To test against an upstream checkout,
+add `-e <path-to-datatrove>` to the install. `docs.bin` is UTF-8 docs joined by `\0`.
+
+Wall time is unreliable on a loaded machine; A/B Rust changes with `tools/icount.sh` (instruction counts are
+load-independent) and confirm wall time by alternating old/new binaries.
+
+After upgrading Python or `regex`, re-export the Unicode tables with `tools/export_pytext.py`.
+
+Rust code is formatted with the repo's `rustfmt.toml`; CI (`.github/workflows/ci.yml`) runs `cargo fmt --check`,
+`clippy -D warnings`, `cargo test --release` and the pytest suite on Python 3.13 against datatrove 0.10.1 and the
+latest 0.10.x.
+
+## Releasing
+
+1. Bump `version` in `pyproject.toml` (the published version; the crate versions don't matter) and merge to `master`.
+2. Tag and push: `git tag v0.1.0 && git push origin v0.1.0`. The tag must equal `v` + the pyproject version.
+3. `.github/workflows/release.yml` builds abi3 wheels (Linux x86_64/aarch64 manylinux_2_28, macOS x86_64/arm64,
+   Windows x64) and an sdist, import-tests each, and publishes to PyPI by trusted publishing.
+
+One-time setup:
+
+- PyPI: on https://pypi.org/manage/account/publishing/ add a pending publisher (or, once the project exists,
+  Manage project > Publishing): PyPI project name `datatrove-rs`, owner `Hexay`, repository `datatrove-rs`, workflow
+  `release.yml`, environment `pypi`.
+- GitHub: Settings > Environments > New environment `pypi` (optionally with required reviewers and a `v*` tag rule).
+