Betlang
CPU source-language detection for code with a tiny 100kb model.
[]
= "0.0.1"
let detection = detect;
assert_eq!;
Use betlang::detect(source) for UTF-8 source strings. Use
betlang::detect_bytes(bytes) for scanners that already work with file bytes
and should not reject non-UTF-8 input before classification. Both return a
Detection; call Detection::language() to read the top language.
Call Detection::top_languages() when you need ranked probabilities.
Supported Languages
Slugs parse through the standard FromStr implementation:
assert_eq!;
asm, awk, batch, bash, c, c-sharp, clojure, cmake, cobol,
commonlisp, cpp, css, dart, diff, dockerfile, elixir, erlang,
go, groovy, haskell, hcl, html, ini, java, javascript,
jinja2, json, julia, kotlin, lua, markdown, matlab, objc,
ocaml, perl, php, postscript, powershell, prolog, python, r,
ruby, rust, scala, scss, solidity, sql, starlark, swift,
textproto, toml, typescript, vb, verilog, vhdl, vue, xml,
yaml, zig.
Several embedded model classes intentionally map to one public language. For
example, erb, gemfile, and gemspec map to ruby; jsonl maps to json;
shell maps to bash; and project-file classes such as csproj and vcxproj
map to xml.
Model
The embedded model is assets/magika/source-student-q4.bin, a 100,444-byte
raw tensor payload with SHA-256:
e2498dc23a60cc32ae21a448c3763ee7080a6fbf9f813b63a066ef195e1e44a0
Architecture: wordseq-b1536-k3-m2048-med-3conv-hidden, tokenizer version 3.
On the held-out bigorig test split it reaches
test_teacher_parity=0.967618 and test_fs_accuracy=0.962517.
See MODEL_CARD.md for the training and evaluation summary.
Performance
Betlang uses a fixed 4096-byte Magika window and pads runtime inference to the
same 2048-token shape used by evaluation. The model is loaded once per process
and then reused through a OnceLock.
Native CPU inference dispatches through fearless_simd. Benchmark entry points
are available through cargo bench. Current baseline numbers are tracked in
BENCHMARKS.md.
License And Attribution
Betlang is licensed under MIT. The embedded student model was trained from outputs of Google's Magika teacher model; Magika is published by Google under Apache-2.0. Keep this attribution with redistributed model artifacts.
Confusion By File Size
The shipped wordseq model is evaluated below on the held-out bigorig test
split. Each panel is a row-normalized confusion matrix for one file-size
bucket: actual labels are rows, predicted labels are columns, and the diagonal
is correct classification.
