Atlas

Models

← All models

Llama 3.1 8B

Meta · 2024-07-23 · 8B parameters

Llama 3.1 8B is Meta's pretrained 8B-class base language model from the July 2024 Llama 3.1 release. It is a dense, multilingual, decoder-only Transformer with a checkpoint-configured 131,072-token context window, grouped-query attention, and a tiktoken-based tokenizer; Meta reports about 15 trillion pretraining tokens and a December 2023 data cutoff for the family. It is intended as a foundation for fine-tuning and downstream natural-language generation rather than as the instruction-tuned chat variant.

Benchmark scores

BenchmarkScore
BBH (Open LLM Leaderboard v2)46.6
Capability103.1
Epoch Capabilities Index (ECI)115
Global PIQA - Non-Parallel (Log-Likelihood Accuracy)56.9
Global PIQA - Parallel (Log-Likelihood Accuracy)25.2
GPQA (Open LLM Leaderboard v2)31.0
IFEval (Open LLM Leaderboard v2)12.5
MATH Level 5 (Open LLM Leaderboard v2)6.6
MEDIC (closed-ended)50.1
MMLU-Pro (Open LLM Leaderboard v2)32.9
MuSR (Open LLM Leaderboard v2)38.1
Loading Atlas data…