Atlas

Models

← All models

Stable Beluga 2

Stability AI · 2023-07-20 · 69B parameters

Stable Beluga 2 is Stability AI's instruction-tuned derivative of Meta's Llama 2 70B, fine-tuned using Stability AI’s internal Orca-style synthetic dataset derived from conceptofmind datasets created by Enrico Shippole to improve multi-step reasoning and instruction-following quality. Released in July 2023, it was among the strongest openly available 70B chat models at the time, demonstrating that Orca-style training could substantially enhance the capability of Llama 2 base models. The model is available on Hugging Face under Stability AI's separate non-commercial community license.

Benchmark scores

BenchmarkScore
ARC AI286.1
BBH (BIG-Bench Hard)69.3
BBH (Open LLM Leaderboard v2)58.2
BoolQ89.4
Capability110.8
Epoch Capabilities Index (ECI)116
GPQA (Open LLM Leaderboard v2)31.6
GSM8K69.6
HellaSwag (Unspecified Scoring Protocol)84.1
IFEval (Open LLM Leaderboard v2)37.9
LAMBADA71.3
MATH Level 5 (Open LLM Leaderboard v2)4.4
MMLU68.6
MMLU-Pro (Open LLM Leaderboard v2)33.3
MuSR (Open LLM Leaderboard v2)47.3
PIQA83.3
Loading Atlas data…