Atlas

Models

← All models

Llama 2 7B Chat

Meta · 2023-07-18 · 6.7B parameters

Llama 2 7B Chat is Meta's smallest Llama 2-Chat checkpoint, a text-only autoregressive language model optimized for dialogue. It was initialized from Llama 2 7B and aligned through supervised fine-tuning followed by iterative RLHF using rejection sampling and PPO; the released artifact supports a 4,096-token context window.

Benchmark scores

BenchmarkScore
AlpacaEval 2.0 (length-controlled win rate)5.4
AlpacaEval 2.0 (raw win rate)5.0
Artificial Analysis Intelligence Index4.3
BBH (Open LLM Leaderboard v2)31.1
BRIDGE Medical (chain-of-thought)13.7
BRIDGE Medical (few-shot)22.3
BRIDGE Medical (zero-shot)16.5
Capability85.7
Creative Writing v246.7
EQ-Bench v236.3
EQ-Bench v2 + MAGI-Hard Combined31.9
GPQA (Open LLM Leaderboard v2)25.3
GPQA Diamond22.7
Humanity's Last Exam (Text-Only)5.8
IFEval (Open LLM Leaderboard v2)39.9
LiveCodeBench (Artificial Analysis)0.2
MAGI-Hard27.5
MATH Level 5 (Open LLM Leaderboard v2)2.0
MATH-500 (Artificial Analysis source)5.9
MMLU Pro16.4
MMLU-Pro (Open LLM Leaderboard v2)16.9
MuSR (Open LLM Leaderboard v2)36.8
SciCode (Artificial Analysis)0.0
Loading Atlas data…