Muose-100M-Decision

Muose-100M-Decision is a 100M parameter decision model developed by Muose.

It is built on Muose-100M, a custom Transformer pretrained from random initialization on 10B tokens and post-trained for probabilistic decision tasks.

Capabilities

• Choice
• Noul
• Score

Results

• Muose-100M-Decision: 70.80% accuracy
• Muose-50M-Decision: 55.25% accuracy

Evaluation was performed on the same 2,000-example mixed evaluation set.

Architecture

• 100M parameters
• MuoseDecision
• Custom Muose Transformer
• Custom 24k ByteLevel BPE tokenizer
• 512 token context
• PyTorch
• Safetensors

Usage

Install dependencies with pip install -r requirements.txt, edit the example state, instructions, and criteria in run.py, then run:

python run.py

Limitations

This is an experimental small model, and performance can vary on unseen domains.

License

CC BY-NC-SA 4.0