Dataset · TIGER-Lab
Programming tasks with synthetic test cases
TIGER-Lab/AceCode-87K
Coding task dataset with automatically generated and filtered test cases for training models via RL.
What's inside
English text dataset containing between 10k and 100k examples (87k in the title). It is built on a base set of programming tasks: powerful LLMs generated verification test cases for them, after which noisy and incorrect variants were filtered out. The data was created specifically for training reward models and reinforcement learning (RL) in code generation.
What you can do with it
- Training reward models to evaluate the correctness of generated code
- RL fine-tuning of coding LLMs with solution verification using automated tests
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 24.2K
- Likes
- 57
Models trained on it
- mradermacher/marin-8b-instruct-i1-GGUFText generation 1.4K
- mradermacher/AceCoder-Qwen2.5-Coder-7B-Base-Rule-i1-GGUF 1.3K
- mradermacher/AceCoder-Qwen2.5-7B-Ins-Rule-i1-GGUF 1.1K
- marin-community/marin-8b-instructText generation 1K
- mradermacher/AceCoder-Qwen2.5-Coder-7B-Base-Rule-GGUF 731
- mradermacher/marin-8b-instruct-GGUFText generation 407
- mradermacher/AceCoder-Qwen2.5-7B-Ins-Rule-GGUF 350
- TULLUS/marin-8b-instruct-GGUFText generation 315
How to load it
from datasets import load_dataset
ds = load_dataset("TIGER-Lab/AceCode-87K")