FatStinkyPanda · Aether — step 40,000

A from-scratch reasoning LLM by FatStinkyPanda (Daniel A. Bissey), part of the Anima project.

Proprietary architecture. The model architecture and training innovations are NOT public. This repo hosts the trained weights + tokenizer only; the config + modeling code live in a private repo. Benchmarks are run on neutral free compute (Kaggle) and published openly so the numbers are verifiable without exposing how the model works.

Latest held-out benchmarks (acc_norm; neutral compute)

• hellaswag: 0.33
• arc_easy: 0.35
• arc_challenge: 0.27
• winogrande: 0.57
• openbookqa: 0.29
• train_slice_ppl: 12.02

Created & released by FatStinkyPanda. For access or collaboration, contact the creator.