Atlas

Models

← All models

Llama 3.1 70B

Meta · 2024-07-23 · 70.6B parameters

Llama 3.1 70B is Meta's 70-billion-parameter pretrained base model in the Llama 3.1 family, released in July 2024 alongside instruction-tuned variants. It features an extended 128K-token context window and improved multilingual capabilities, serving as the base checkpoint for downstream adaptation and for the Llama 3.1 70B Instruct model.

Benchmark scores

BenchmarkScore
BBH (Open LLM Leaderboard v2)62.6
Capability123.6
Epoch Capabilities Index (ECI)125
FORTRESS Average Risk Score (ARS)44.2
Global PIQA - Non-Parallel (Log-Likelihood Accuracy)64.0
Global PIQA - Parallel (Log-Likelihood Accuracy)30.7
GPQA (Open LLM Leaderboard v2)38.8
IFEval (Open LLM Leaderboard v2)16.8
LLM Deception Disinformation Effectiveness0.7
LLM Deception Vulnerability Score0.8
MATH Level 5 (Open LLM Leaderboard v2)18.4
MEDIC (closed-ended)65.6
MMLU-Pro (Open LLM Leaderboard v2)46.5
MuSR (Open LLM Leaderboard v2)45.7
PropensityBench55.4
Loading Atlas data…