---
tags:
  - pre-pretraining
  - shuffle_dyck
---

# PPT Corpus: `shuffle_dyck`

Raw stage-1 pre-pretraining corpus from the [ppt](https://github.com/sashaboguraev/ppt)
research framework, exported for information-theoretic analysis (e.g.
m-local entropy) independent of this repo's training pipeline.

## IMPORTANT: token IDs are NOT standard BPE tokens

Raw Shuffle-Dyck bracket-symbol IDs (integer `i` = opening bracket type `i`, `i+k` = its closing counterpart), generated by `submodules/pre-pretraining/src/grammar.py::generate_shuff_dyck`.

**These `input_ids` are not decodable with the Pythia (or any other)
tokenizer.** They are integers in `[0, 128)` with regimen-specific
meaning (see above) — treat this dataset as a corpus of symbol/event
sequences, not text.

## Corpus stats

| Field | Value |
|-------|-------|
| Vocabulary size | 128 |
| Chunk length (`seq_len`) | 2048 |
| Number of chunks | 50,000 |
| Total tokens | 102,400,000 |
| Generation seed | None |

Each row's `input_ids` is a fixed-length list of 2048 integers — the
exact chunk granularity fed to the model during stage-1 training (formed
by concatenating generated sequences and splitting into non-overlapping
2048-token blocks).



## Generation config

```json
{
  "num_symbols": 64,
  "n": 200000,
  "target_length": 512,
  "min_depth": 1,
  "max_depth": 8,
  "raw_ids": true,
  "file_dir": "data/raw/shuffle_dyck",
  "cache_dir": "data/tokenized/shuffle_dyck_raw",
  "seq_len": 2048
}
```

## Usage

```python
from datasets import load_dataset

ds = load_dataset("sashaboguraev/ppt-shuffle_dyck-corpus")
example = ds["train"][0]["input_ids"]  # list[int], length 2048
```