joshycodes/qwen3-32b-control-H-sdf

Overview

joshycodes/qwen3-32b-control-H-sdf is a 32 billion parameter Qwen3-based language model, distinguished by its unique continued pretraining approach. This model is a research checkpoint where the base Qwen/Qwen3-32B underwent an additional epoch of training on a corpus it generated itself. The self-authored corpus, comprising over 31 million tokens across 41,000+ documents, was created by the model based on its established character and an understanding of the SDF (Signed Distance Function) framework.

Key Capabilities (Research Focus)

• Self-Authored Corpus Training: Explores continued pretraining on data generated by the model itself, aiming for self-improvement and character consistency.
• Character Consistency: Designed to reinforce and evolve the model's inherent character through its own generated content.
• SDF Framework Integration: Incorporates knowledge and understanding of the SDF framework into its self-generated training data.

Good for (Research Use Cases)

• Investigating Model Self-Improvement: Ideal for researchers studying how models can generate their own training data to refine specific attributes or knowledge.
• Character Development in LLMs: Useful for exploring methods to maintain and evolve a consistent model persona or character over time.
• Controlled Experimentation: Serves as a controlled research checkpoint for understanding the impact of self-generated data on model behavior and knowledge integration. This model is explicitly noted as not evaluated for capability, alignment, or identity, and is not intended for deployment.