AresXLG1

AresXLG1 is the 100 million parameter model in the Ares model family, it achieves top-tier performance across many benchmarks and on the Open SLM Leaderboard.

It is trained on general, heavily filtered data, from filtered DCLM, UltraFineWeb L3, PerfectPdfs, the smollm corpus, MegaMath-Web-Pro-Max, and more.

It also uses many private datasets which include data distilled from larger models and other unfinished high-quality datasets and some that will never be public.

Benchmarks

AresXLG1-100M:

ArithMark 3.0 PIQA HellaSwag ARC-Easy ARC-Challenge Open SLM Int Index
60.60 68.25 46.58 58.00 32.08 34.84

KLM 1.6:

ArithMark 3.0 PIQA HellaSwag ARC-Easy ARC-Challenge Open SLM Int Index
55.90 70.18 45.28 61.07 33.53 33.95

Boris 2 (144M):

ArithMark 3.0 PIQA HellaSwag ARC-Easy ARC-Challenge Open SLM Int Index
45.90 72.85 50.27 59.76 32.00 34.34

Boris 2 Think (144M):

ArithMark 3.0 PIQA HellaSwag ARC-Easy ARC-Challenge Open SLM Int Index
49.20 72.31 50.13 63.93 34.81 36.05

Architecture And Tokenizer

It uses an architecture and tokenizer heavily inspired by the latest KLM models while outperforming them in several areas.

Training Info

4x B300 GPUs, 50 billion tokens.

Warning

This is the base model with no instruct or thinking tuning, it cannot be used to chat, think, or follow instructions.

Downloads last month
-
Safetensors
Model size
99M params
Tensor type
I64
·
F32
·
I32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Datasets used to train ArrowLabs/AresXLG1-100M