Dataset · argilla
Synthetic instructions for checking strict format adherence
argilla/ifeval-like-data
550 thousand instruction-response pairs in the style of the IFEval benchmark from Argilla for training models to precisely follow constraints.
What's inside
The dataset from Argilla contains between 100 thousand and 1 million examples in English (550 thousand rows in the base subset) synthetically generated by the Qwen2.5-72B-Instruct model using the MagPie methodology. Prompts are constructed by analogy with the google/IFEval benchmark, may include conflicting requirements, and answer correctness was verified via lm-evaluation-harness.
What you can do with it
- SFT fine-tuning of models for strict compliance with complex and formal instructions
- Improving LLM performance on the IFEval benchmark and similar controllability tests
- Studying model robustness to conflicting prompt requirements
- Tasks
- Text generation
- Languages
- English
- Size
- 100K–1M rows
- Downloads this month
- 5.8K
- Likes
- 51
Models trained on it
- mradermacher/Unbound-v1.12.0-27B-i1-GGUF 1.1K
- mradermacher/Mira-v1.12.1-27B-i1-GGUF 846
- mradermacher/Unbound-v1.12.0-27B-GGUF 717
- mradermacher/Mira-v1.12.1-27B-GGUF 545
- Diamegs/PIT-4B-FT-201812Text generation 149
- Diamegs/PIT-4B-FT-202012Text generation 142
- Diamegs/PIT-4B-FT-202412Text generation 134
- Diamegs/PIT-4B-FT-202212Text generation 128
How to load it
from datasets import load_dataset
ds = load_dataset("argilla/ifeval-like-data")