---
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
  - code
pretty_name: DragonCode Tokenized Pretrain (Domain-Code)
size_categories:
  - 10B<n<100B
---

# DragonCode Tokenized Pretrain — Domain-Code Corpus

Pre-tokenized training corpus for the DragonCode LLM family. This is the
**domain-code-only** dataset (not general web text).

## Format

- 26 shards: `data/train-00000-of-00026.bin` .. `train-00025-of-00026.bin`
- Flat `int32` little-endian token-id stream (no headers, no row structure).
  Concatenate shards in order to reconstruct the full sequence.
- **3,417,141,141 tokens** total (~3.42B), ~13.67 GB on disk.
- Sequence length: 4096 (`MAX_SEQ_LEN`) during training.

## Sources (streaming, permissive only)

| Source | Field | License filter |
|---|---|---|
| `codeparrot/codeparrot-clean` | `content` | permissive only (MIT/ISC/Apache-2.0/BSD/CC0/Unlicense/etc.) |
| `open-r1/codeforces-cots` | `prompt` / `generation` | permissive only |

Non-permissive-licensed and empty samples are dropped. General web corpora
(FineWeb-Edu / SlimPajama / The-Pile) are intentionally excluded per project spec.

## Tokenizer

Uses the DragonCode tokenizer (see `DragonLimited/DragonCode-150M`). Encode with
`add_special_tokens=False`.

## Note

The previous 5-shard legacy corpus (FineWeb-Edu-based) was fully replaced by this
domain-code corpus on 2026-08-25.