Quirky Models research suite
Quirky Models deliberately trains language models to give systematic wrong answers under a named condition. The research then studies whether probes can recover the information the model still represents internally.
Research release
A controlled reason for incorrect answers
The project uses a Bob condition that triggers unreliable behavior and an Alice condition associated with truthful answers. The repository describes 96 LoRA adaptations across 12 tasks. This is a research collection, not a general-purpose chat product.Quirky Models and eliciting latent knowledge research
Quirky Math changes the prompt arrangement
The Quirky Math subset contains 24 models with mixture, grader-first and grader-last prompt layouts. The Pythia 410M mixture checkpoint is one documented example. Those prompt variants belong to the experiment rather than a ranking of general model quality.Quirky Math Pythia 410M model card
Base-model terms still apply
The collection includes adaptations of several Pythia sizes, Llama 2 7B and Mistral 7B. Architecture, memory requirements and license terms therefore depend on the selected base. Do not apply a license label from one checkpoint to the entire mixed collection.Quirky Models and eliciting latent knowledge researchOfficial Quirky Models and datasets collection
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | EleutherAI/qm-pythia-410m-mixture |
| Adaptation method | LoRA fine-tuningQuirky Models and eliciting latent knowledge research |
| Collection scope | 96 adaptations across 12 tasks, as described by the projectQuirky Models and eliciting latent knowledge research |
| License scope | Varies with the selected checkpoint and underlying base modelOfficial Quirky Models and datasets collection |
Related models
- Pythia 1B
EleutherAI · Pythia
A wide, sixteen-layer Pythia model trained with bfloat16 precision.
Launch date Research releaseDownloadable - Pythia 410M
EleutherAI · Pythia
A 24-layer research model with a 1,024-wide hidden state.
Launch date Research releaseDownloadable