Dataset · TIGER-Lab
Answer critique instructions for LLM training
TIGER-Lab/WebInstruct-CFT
An English dataset of 100K-1M examples with error breakdowns in responses for training models via critique.
What's inside
An English text data collection ranging from 100k to 1 million records, created based on WebInstruct. Unlike standard SFT datasets with reference answers, it contains critiques of ready-made responses. The dataset was released by the TIGER-Lab team along with a research paper and trained models of 7B and 32B parameters.
What you can do with it
- Fine-tuning LLM using Critique Fine-Tuning (CFT) to find reasoning errors
- Training critic models and verifiers to evaluate generations
- Tasks
- Text generation
- Languages
- English
- Size
- 100K–1M rows
- Downloads this month
- 2.2K
- Likes
- 60
Models trained on it
- mradermacher/Qwen2.5-32B-Instruct-CFT-i1-GGUF 1.2K
- bartowski/TIGER-Lab_Qwen2.5-32B-Instruct-CFT-GGUFText generation 1.1K
- mradermacher/Qwen2.5-Math-7B-CFT-i1-GGUF 908
- mradermacher/Qwen2.5-32B-Instruct-CFT-GGUF 345
- mradermacher/Qwen2.5-Math-7B-CFT-GGUF 171
- tensorblock/Qwen2.5-32B-Instruct-CFT-GGUFText generation 86
- TIGER-Lab/Qwen2.5-Math-7B-CFTText generation 64
- TIGER-Lab/Qwen2.5-32B-Instruct-CFTText generation 38
How to load it
from datasets import load_dataset
ds = load_dataset("TIGER-Lab/WebInstruct-CFT")