---
pretty_name: lex-large
task_categories:
- token-classification
tags:
- source-code
- syntax-highlighting
---

# lex-large

Prepared numeric training cache for the 101,107-byte, 185-grammar neural lexer.
It contains token features and nine-class syntax labels, not the original source
text. File-level hashing fixes the train/validation/test split before upload.

## Contents

- `train.npz`, `val.npz`, `test.npz`: ragged token windows and split indices
- `meta.json`: counts, label distribution, and per-language coverage
- `language_manifest.json`: pinned Highlight.js grammar mapping
- `sources.json`: GitHub repository provenance used by corpus preparation
- `gpu_lexer_import.json`: hashes and counts for the imported gpu-lexer
  train/mining shards; the verification split is excluded
- `checksums.json`: byte sizes and SHA-256 checksums

Total labelled tokens: 136,430,550

Languages covered: 185/185

Sequence length: 512

## Download

```bash
cd /path/to/lexer/train_large
hf download anudit/lex-large-stack-corpus --repo-type dataset --local-dir ./corpus/dataset_v2
.venv/bin/python test_setup.py
```

The cache is consumed directly by `train_teacher.py`, `train.py`, and
`eval_coverage.py`. See `train_large/note.txt` in the lexer repository for the
complete commands.

## Integrity

This revision contains 7 payload files. Verify them against
`checksums.json` after download.