swadeshb/comem-qwen3-4b-regret-gate

CoMem Qwen3-4B Regret-Gate Manager

This model is a merged, directly loadable checkpoint from an experimental 310-trajectory regret-weighted CoMem manager. It is based on YWZBrandon/summary-sft-qwen3-4b and features a 4 billion parameter Qwen3 architecture.

Key Capabilities & Training:

• Summary Generation: Trained with GRPO (Generalized Reinforcement Learning with Policy Optimization) using DeepSWE next-action preservation.
• Bounded Context Length Advantages: Optimized for scenarios with limited token budgets.
• KEEP/SUM Selector: Utilizes regret-weighted soft targets and action-balanced replay for intelligent content selection.
• Context Handling: Manages eight summaries when above an 8,192-token budget, and one virtual KEEP reference with seven summaries at or below budget.

Experimental Caveats:

• Gate Evaluation: The current one-token, unrestricted text-generation gate evaluation may emit a third token instead of the intended KEEP or SUM verbalizers. For accurate gate measurement, it is recommended to compare next-token logits of KEEP and SUM or use constrained decoding.
• Validation Status: This checkpoint is experimental and has not yet undergone full validation via a complete SWE-bench Verified harness run.

Good For:

• Experimental research in regret-weighted content management and summarization.
• Applications requiring intelligent selection and summarization of text within specific token constraints.
• Developers interested in exploring advanced reinforcement learning techniques for language model control.