Spaces:

---

AxiomicLabs

/

Open_SLM_Leaderboard

like 65

Running

Submitting Nexus-Erebus-50M by Ideoa Labs for the ~50M class.

Model (open weights, Apache 2.0): https://huggingface.co/MaliosDark/Nexus-Erebus-50M
Trained from scratch (random init, own digit-atomic tokenizer, no base model).

0-shot results, full test sets (acc_norm for MC, accuracy for ArithMark-2):

Evaluated with lm-eval-harness (no subsampling) and the public Arithmark-2.0 script.

Method note, for transparency: the model is trained on the official train splits of the
public benchmarks plus synthetic integer arithmetic (addition, subtraction, multiplication,
exact division, mixed and parenthesised multi-operator expressions). No evaluation items
were used at any stage.