Add Nexus-Erebus-3M (from-scratch 3.0M, arithmetic specialist)
#40
by MaliosDark - opened
Submitting Nexus-Erebus-3M for the small-model class.
Model (open weights, MIT): https://huggingface.co/MaliosDark/Nexus-Erebus-3M
Trained 100% from scratch: random init, own digit-atomic tokenizer, no base model.
3.0M parameters (3,000,384). An integer-arithmetic specialist.
0-shot, full test sets (acc_norm for MC, accuracy for ArithMark-2):
| Task | Score |
|---|---|
| ArithMark-2 | 70.72 |
| PIQA | 51.41 |
| HellaSwag | 26.59 |
| ARC-Easy | 27.02 |
| ARC-Challenge | 20.65 |
The tokenizer applies a least-significant-digit-first transform internally, so the model must be
loaded with trust_remote_code=True (same pattern as other specialist entries). A self-contained
verification script is included in the repo:
python benchmark_nexus_arithmark.py MaliosDark/Nexus-Erebus-3M
Happy to provide anything needed for independent verification.
Datdanboi25 changed pull request status to merged
Shes up!