Deep Ignorance research models
Deep Ignorance compares a baseline and altered training runs to study whether excluding selected information can reduce harmful capabilities. The released study includes 18 models at a 6.9B scale.
Research release
A family of interventions with a shared baseline
The release includes an unfiltered baseline, strong filtering throughout training, and a run that changes the filter during the final training stage. It also includes a post-trained baseline. These are research interventions, not different sizes or commercial service tiers.Deep Ignorance model cardDeep Ignorance research code and model variants
Inspect the run as well as its final model
The model card lists 101 pretraining checkpoints and 11 final-stage checkpoints. The project evaluates whether interventions remain effective after further training. Its findings concern the studied settings; they do not guarantee that a downloadable model is safe in every application.Deep Ignorance model card
Keep the experiment name with the weights
Practical guidance. Select a named variant from the project repository and record its training-step revision. The suite is grouped here because the variants are controlled experiments within one research program.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Launch date | August 2025 |
| Access | Downloadable |
| Model identifiers | EleutherAI/deep-ignorance-e2e-strong-filter |
| Scale per model | 6.9 billion parametersDeep Ignorance model card |
| Published study | 18 model variantsDeep Ignorance model card |
| License | Apache 2.0Deep Ignorance model card |
Related models
- Pythia 6.9B
EleutherAI ยท Pythia
A 32-layer open model for research at a six-billion-parameter scale.
Launch date Research releaseDownloadable