Atlas

Benchmarks

← All benchmarks

Needle-in-a-Haystack multi-needle (Meta protocol)

General QA · 2024-07-23

Meta's multi-needle variant of Needle-in-a-Haystack used for Llama 3.1 and later: four needles are inserted into a long document and the model must retrieve two of them, at context lengths evenly spaced from 2,000 to 131,072 tokens (128,000 for non-Llama models) in ten steps. Scores are average recall across the lengths. It is a synthetic retrieval probe close to saturation and is kept as a diagnostic.

Top models (higher is better)

ModelScore
GPT-4 0125 Preview100.0
GPT-4o (source unspecified)100.0
Llama 3.1 8B Instruct98.8
Llama 3.1 405B Instruct98.1
Llama 3.1 70B Instruct97.5
Claude 3.5 Sonnet (June 2024)90.8
Llama 3.2 3B Instruct84.7
Llama 3.2 1B Instruct75.0
Phi-3.5 Mini Instruct52.7
Loading Atlas data…