metadata
datasets:
- CEAMFA/rewrite
- HuggingFaceFW/fineweb-edu
- appvoid/no-prompt-15k
language:
- en
tags:
- base
- sml
- void
- pretrained-from-scratch
Introducing void: our first ever language model, trained from scratch with a novel hybrid tokenizer on 300M high-quality tokens (total of 2 epochs on a B300) using 4096 as context window. Total cost was $23 dollars. 132m parameters. Future releases are expected to be published in the following weeks/months.
| Benchmark | Accuracy | Normalized |
|---|---|---|
| ARC Challenge | 25.17% | 27.22% |
| ARC Easy | 45.03% | 43.01% |
| HellaSwag | 31.57% | 35.77% |
| PIQA | 61.15% | 60.12% |
| WinoGrande | 53.12% | — |
| ArithMark | 35.20% | 35.20% |
Disclaimer: Even though the model is based on gemma 3 architecture, the tokenizer is different so you might need to wait until this model can be added to llama.cpp