---
license: odc-by
task_categories:
- text-generation
language:
- en
tags:
- parameter-golf
- bpe
- sentencepiece
- fineweb
---

# parameter-golf-fineweb-sp16384

SentencePiece BPE tokenizer (`vocab_size=16384`) and pre-tokenized
[FineWeb-10B](https://huggingface.co/datasets/HuggingFaceFW/fineweb) shards

## Structure

- **`canonical/`** — used by the submission script. Plain sp16384 BPE
  tokenization of FineWeb-10B (50K validation docs + 80 train shards).
- **`caseops/`** — *not used by the submission*. Experiment artifacts from
  a CaseOps lossless-case-transform variant.

## Submission download

The submission's `setup.sh` pulls only the `canonical/` subdirectory:

```bash
hf download deniskurlov/golf-16k-tokenizer --include "canonical/**" --local-dir ./data --repo-type dataset
```

This skips the ~12 GB CaseOps shards.

## File layout

```
canonical/
├── tokenizers/
│   ├── fineweb_16384_bpe.model
│   └── fineweb_16384_bpe.vocab
└── datasets/
    └── fineweb10B_sp16384/
        ├── fineweb_train_*.bin   (80 shards, ~125 MB each)
        └── fineweb_val_*.bin

caseops/
├── tokenizers/
│   ├── fineweb_16384_bpe_lossless_caps_caseops_v1_reserved.model
│   └── fineweb_16384_bpe_lossless_caps_caseops_v1_reserved.vocab
└── datasets/
    └── fineweb10B_sp16384_lossless_caps_caseops_v1_reserved/
        ├── fineweb_train_*.bin
        ├── fineweb_val_*.bin
        └── fineweb_val_bytes_*.bin    (per-token original-byte counts for BPB)
```