Atlas

Models

← All models

Llama 2 70B Chat

Meta · 2023-07-18 · 69B parameters

Llama 2 70B Chat is Meta's 70-billion-parameter instruction-tuned and RLHF-aligned language model from the Llama 2 family, representing the flagship assistant model of the Llama 2 generation. Fine-tuned with supervised instruction following and iterative reinforcement learning from human feedback, it was positioned as Meta's most capable open-weight conversational model at the time of release. It excels at helpfulness and safety-conscious dialogue across a broad range of tasks.

Benchmark scores

BenchmarkScore
AlpacaEval 2.0 (length-controlled win rate)14.7
AlpacaEval 2.0 (raw win rate)13.9
Artificial Analysis Intelligence Index3.0
BBH (Open LLM Leaderboard v2)30.4
BRIDGE Medical (chain-of-thought)19.0
BRIDGE Medical (few-shot)27.1
BRIDGE Medical (zero-shot)25.1
Capability103.3
EQ-Bench v273.6
EQ-Bench v2 + MAGI-Hard Combined54.5
ForecastBench Baseline Dataset Brier Index50.2
ForecastBench Baseline Market Brier Index51.7
ForecastBench Baseline Overall Brier Index50.9
ForecastBench Preliminary Dataset Brier Index50.2
ForecastBench Tournament Dataset Brier Index50.2
ForecastBench Tournament Market Brier Index52.5
ForecastBench Tournament Overall Brier Index51.4
GPQA (Open LLM Leaderboard v2)26.4
Humanity's Last Exam (Text-Only)5.0
IFEval (Open LLM Leaderboard v2)49.6
LiveCodeBench (Artificial Analysis)9.8
MAGI-Hard35.4
MATH Level 53.3
MATH Level 5 (Open LLM Leaderboard v2)2.9
MATH-500 (Artificial Analysis source)32.3
MMLU Pro40.6
MMLU-Pro (Open LLM Leaderboard v2)24.3
MuSR (Open LLM Leaderboard v2)36.9
OTIS Mock AIME 2024-20250.0
Loading Atlas data…