Nous ResearchOLMo Bitnet

OLMo Bitnet 1B

This 1B model tests the 1.58-bit training method using OLMo software and Dolma data.

Research release

Training and model behavior

Nous trained the proof of concept on the first sixty billion tokens of Dolma. A separate run used standard FP16 weights with the same hyperparameters for comparison. The release includes example inference code.OLMo Bitnet 1B model card

Working with this model

Practical guidance. Use this artifact to study the training method and its implementation. Do not assume a generic quantized-model loader reproduces the training scheme.

Specifications & access

OLMo Bitnet 1B specifications
DetailValue and conditions
StatusResearch release
Launch dateMarch 2024
AccessDownloadable
Model identifiersNousResearch/OLMo-Bitnet-1B
Repository license labelapache-2.0OLMo Bitnet 1B model card
Parameter scale1BOLMo Bitnet 1B model card