fixtures/FIXTURES.md

Multilingual & modality benchmark fixtures

Plain-text fixtures for benchmarking encode performance across languages, scripts, and content types. Built by fetch_fixtures.py (same directory); exact per-file provenance — source dataset, config, split, pinned revision SHA, byte/doc counts, build date — is recorded in fixtures_manifest.json.

Layout:

• lang/ — natural-language corpora, one <lang>_<script>.txt per language
• modalities/ — code, math, and agentic-trace corpora

Note: this whole data/ directory is gitignored. To rebuild any file from scratch:

uv run --python 3.12 --with 'datasets>=3.2' fetch_fixtures.py [name ...]

Natural languages — lang/<lang>_<script>.txt

Named by ISO 639-3 language code + ISO 15924 script code, matching the config names of HuggingFaceFW/fineweb-2, the source for all non-English files (real, deduplicated, filtered web text; ODC-By license). ~5 MB per language, whole documents separated by blank lines, taken in dataset order from the (small) test split at a pinned revision. eng_Latn.txt comes from HuggingFaceFW/fineweb (sample-10BT) as the Latin-script baseline.

Coverage: Devanagari (hin), Bengali (ben), Tamil (tam), Cyrillic (rus), Arabic (arb, RTL), Hebrew (heb, RTL), Japanese (jpn), Hangul (kor), Han (cmn), Thai (tha, no word separators), Greek (ell), Georgian (kat), Ethiopic (amh), Latin (eng).

Modalities — modalities/

• code_mixed.txt — source code, ~0.75 MB each of Python/JS/TS/Rust/Go/ Java/C/C++ from GitHub tarballs of well-known OSS repos (django, jquery, vue, tokio, caddy, junit5, redis, fmt), pinned to the commit SHAs recorded in the manifest and read in memory only. bigcode/the-stack-smol would have been the HF-native choice but is gated and no valid HF token was available at build time.
• math_latex.txt — mathematical web text with LaTeX from open-web-math/open-web-math.
• agentic_swe.txt — real SWE-agent coding-assistant trajectories from SWE-bench/SWE-smith-trajectories, serialized as [role]\ncontent blocks (tool calls, diffs, terminal output included). Complements agentic-traces.txt (pre-existing synthetic Claude-Code-style traces, provenance unrecorded — not managed by fetch_fixtures.py, moved here from data/ for tidiness; referenced by tk-encode's pipeline_benchmark.rs and pipeline_oracle.rs).

Safety notes

• fetch_fixtures.py forces HF_DATASETS_TRUST_REMOTE_CODE=0 and only reads parquet-native datasets — no dataset loading scripts execute.
• Revisions are pinned at fetch time (recorded in the manifest), so a rebuild can't silently pull different upstream content.
• Content is raw web/code text: treat it as untrusted data. Fine to feed to tokenizers; don't execute it, and prefer less over cat when eyeballing it (raw web text can contain terminal escape sequences). NUL bytes are stripped at build time; everything else is kept as-is on purpose — the mess is representative.