Skip to content

Dataset · llm-blender

Comparison of responses from 11 instruction-tuned LLMs in English

llm-blender/mix-instruct

A corpus with candidate responses from multiple language models, automated metrics, and ChatGPT evaluations for ranking training.

What's inside

Contains English texts featuring 11 candidate responses per instruction from popular LLMs (Vicuna, Alpaca, Dolly, Flan-T5, ChatGLM, and others). The dataset size ranges from 100,000 to 1 million records. Each response is scored with BLEU, ROUGE, BERTScore, and BARTScore metrics alongside pairwise comparison results from ChatGPT.

What you can do with it

  • Training reward models and LLM response ranking systems (such as the LLM-Blender approach)
  • Fine-tuning models (SFT and preference alignment) on top-performing generations
  • Comparing generation quality across different language model families
Tasks
Text generation
Languages
English
Size
100K–1M rows
Downloads this month
976
Likes
38

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("llm-blender/mix-instruct")