SPP - a model-raising Collection

model-raising 's Collections

Pretraining Datasets

Post-training Dataset

SPP

updated about 3 hours ago

Synthetic Persona Pretraining (SPP): 3B models from the paper.

Models - 3B

Collection

3B SPP models: bases publish their training trajectory as step-* revisions, plus instruct counterparts. • 10 items • Updated 37 minutes ago
Models - 1.7B

Collection

1.7B SPP models: bases publish their training trajectory as step-* revisions, plus instruct counterparts. • 10 items • Updated about 3 hours ago
Pretraining Datasets

Collection

The SPP pretraining corpus: reflections, the corpus selection manifest, safety scores, and verification data. Fully reproducible from public Dolma 3. • 6 items • Updated about 3 hours ago
Post-training Dataset

Collection

SP-SFT: the post-training mix that binds the assistant identity to the persona installed during pretraining. • 2 items • Updated about 3 hours ago
Evals

Collection

ConstitutionEval (in-domain constitution following) and audited AIRiskDilemmas (out-of-distribution value prioritisation and risk-taking). • 2 items • Updated about 3 hours ago