Dataset · tokyotech-llm
Large-scale SwallowCode-v2 code corpus with hundreds of millions of lines
tokyotech-llm/swallow-code-v2
A massive updated code dataset from TokyoTech-LLM for training program generation models.
What's inside
A large text and tabular dataset of 100 million to 1 billion records in English created by the TokyoTech-LLM team (accompanied by arXiv:2505.02881). The second version builds on the SwallowCode-v1 approach (rewriting Python code using LLMs) and overcomes the size and license limitations of the first version.
What you can do with it
- Pretrain and fine-tune language models for code writing
- Research methods for improving code dataset quality through LLM rewriting
- Tasks
- Text generation
- Languages
- English
- Size
- 100M–1B rows
- Downloads this month
- 19.3K
- Likes
- 63
Models trained on it
- mradermacher/GPT-OSS-Swallow-120B-RL-v0.1-i1-GGUF 5.5K
- tokyotech-llm/Qwen3-Swallow-8B-RL-v0.2Text generation 5.3K
- tokyotech-llm/Qwen3-Swallow-8B-SFT-v0.2Text generation 3.2K
- mradermacher/Qwen3-Swallow-30B-A3B-RL-v0.2-heretic-i1-GGUF 1.3K
- mradermacher/GPT-OSS-Swallow-20B-SFT-v0.1-i1-GGUF 1.2K
- mradermacher/GPT-OSS-Swallow-20B-RL-v0.1-i1-GGUF 1.2K
- tokyotech-llm/Qwen3-Swallow-8B-CPT-v0.2Text generation 1.2K
- mradermacher/Unbound-v1.12.0-27B-i1-GGUF 1.1K
How to load it
from datasets import load_dataset
ds = load_dataset("tokyotech-llm/swallow-code-v2")