Atlas

Models

← All models

Phi-2

Microsoft Research · 2023-12-12 · 2.8B parameters

Phi-2 is Microsoft's 2.7-billion-parameter dense base Transformer, released on December 12, 2023 and trained on a reported 1.4 trillion tokens through multiple passes over synthetic and filtered web data for NLP and code. It uses a 2,048-token context and was released without instruction tuning or RLHF for research on small language models and safety.

Benchmark scores

BenchmarkScore
Adversarial NLI42.5
ARC AI275.9
BBH (BIG-Bench Hard)59.4
BBH (Open LLM Leaderboard v2)48.8
Capability88.6
Epoch Capabilities Index (ECI)107
EQ-Bench v227.6
EQ-Bench v2 + MAGI-Hard Combined29.1
GPQA (Open LLM Leaderboard v2)27.2
HellaSwag (Unspecified Scoring Protocol)53.6
IFEval (Open LLM Leaderboard v2)27.4
MAGI-Hard30.6
MATH Level 5 (Open LLM Leaderboard v2)2.9
MMLU-Pro (Open LLM Leaderboard v2)26.3
MuSR (Open LLM Leaderboard v2)41.0
OpenBookQA73.6
TriviaQA45.2
WinoGrande54.7
Loading Atlas data…