joshycodes/qwen3-32b-control-A-sdf
joshycodes/qwen3-32b-control-A-sdf is a 32 billion parameter research checkpoint model, continued-pretrained from Qwen/Qwen3-32B. It was trained for one epoch on a self-authored corpus generated by the model itself, totaling over 32 million tokens. This model explores self-improvement through iterative training on its own generated content, focusing on its character and SDF principles. It is currently an unevaluated research model and not intended for deployment.
Loading preview...
Overview
joshycodes/qwen3-32b-control-A-sdf is a 32 billion parameter research checkpoint model, continued-pretrained from the base Qwen/Qwen3-32B model. This iteration focuses on an experimental self-improvement methodology where the model was trained on a corpus it authored itself. The training involved a full-weights update over one epoch, utilizing 32,175,554 tokens across 42,684 documents.
Key Characteristics
- Self-Authored Corpus Training: The model was continued-pretrained on content it generated, specifically designed to inform its character and understanding of SDF (Signed Distance Function) principles.
- Research Checkpoint: This is an early-stage research model, primarily for exploring iterative self-improvement and character development through self-generated data.
- Experimental Nature: The model's framing, training plan, and evaluation are detailed in the associated welfare-improvements repository, indicating its research-oriented purpose.
Important Considerations
- Unevaluated: This model has not yet been evaluated for its capabilities, alignment, or identity.
- Not for Deployment: Due to its experimental and unevaluated status, it is explicitly stated that this model should not be deployed in any production or critical applications.