InstructEval is a holistic evaluation suite for instruction tuned language models. It exists because a single leaderboard number hides what instruction tuning did and did not buy.

Three axes measured

  • problem solving
  • writing ability
  • alignment to human values

Inference speed is not one of them.

Headline findings

  • instruction-data quality matters more than quantity
  • a model's score is attributed to three causes: the pretraining foundation, the instruction data and the training method
  • open source instruction tuned models are fluent, and fluency is not a good proxy for capability, because they lag on problem solving and alignment

The training method it evaluates is Instruction Tuning.