Atlas

Models

← All models

Phi-1.5

Microsoft Research · 2023-09-11 · 1.4B parameters

Phi-1.5 is Microsoft's open-weight 1.3B-class dense base Transformer, released on September 11, 2023 without instruction tuning or RLHF. Microsoft trained it from random initialization for a reported 150B tokens using a 30B-token dataset composed almost entirely of synthetic textbook-like NLP and coding data; the only non-synthetic component was 6B tokens of filtered code data, not generic web-crawl text. Its 24-layer, 32-head architecture uses a 2,048-token context, partial rotary embeddings over half of each attention head, the CodeGen-Mono tokenizer, Adam, and fp16 training.

Benchmark scores

BenchmarkScore
ARC AI244.4
BBH (Open LLM Leaderboard v2)33.6
BoolQ75.8
Capability84.3
Epoch Capabilities Index (ECI)89.9
GPQA (Open LLM Leaderboard v2)26.8
HellaSwag (Unspecified Scoring Protocol)47.6
IFEval (Open LLM Leaderboard v2)20.3
MATH Level 5 (Open LLM Leaderboard v2)1.8
MedMCQA (Open Medical-LLM)34.3
MedQA 4-options (Open Medical-LLM)31.1
MMLU37.6
MMLU Anatomy (Open Medical-LLM)39.3
MMLU Clinical Knowledge (Open Medical-LLM)46.0
MMLU College Biology (Open Medical-LLM)34.0
MMLU College Medicine (Open Medical-LLM)37.6
MMLU Medical Genetics (Open Medical-LLM)42.0
MMLU Professional Medicine (Open Medical-LLM)29.0
MMLU-Pro (Open LLM Leaderboard v2)16.9
MuSR (Open LLM Leaderboard v2)34.0
OpenBookQA37.2
PubMedQA (Open Medical-LLM)67.8
WinoGrande73.4
Loading Atlas data…