swadeshb/noop-summary-grpo-310-conditional

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 7, 2026Architecture:Transformer Featherless Exclusive Cold

The swadeshb/noop-summary-grpo-310-conditional model is a 4 billion parameter Qwen3-based conditional context manager, fine-tuned from YWZBrandon/summary-sft-qwen3-4b. It was trained using TRL GRPO on 310 software-engineering agent trajectories to implement a hierarchical context-management policy. This model excels at dynamically managing context within an 8,192-token budget by either summarizing older interaction history or retaining full context based on a learned gate. Its primary use is for efficient context handling in agent-based systems, particularly in software engineering domains.

Loading preview...

Model Overview

The swadeshb/noop-summary-grpo-310-conditional is a 4 billion parameter Qwen3-based model designed for conditional context management. Fine-tuned from YWZBrandon/summary-sft-qwen3-4b using TRL GRPO, it was trained on 310 preserved software-engineering agent trajectories.

Key Capabilities

  • Hierarchical Context Management: Implements a policy to manage context within a strict 8,192-token budget.
  • Conditional Summarization: Deterministically summarizes (SUM) older interaction history when above budget, or uses a learned gate to KEEP or SUM when at or below budget.
  • System Prompt Retention: When summarizing, it retains the system prompt, task, and the two most recent interaction messages.
  • Optimized Policy: The summary policy is optimized with GRPO, and the below-budget gate learns from measured counterfactual utility margins.

Training Details

The model was trained for one epoch, utilizing a manager prompt limit of 16,384 tokens and a maximum completion length of 2,048 tokens. It uses a learning rate of 1e-6 and a TRL KL coefficient of 0.01.

Limitations

  • Reward measures preservation of recorded next DeepSWE action, not long-horizon task success.
  • The learned gate is only active at or below the strict budget boundary.
  • Training set is small and specialized, requiring evaluation on target agent and context distributions before deployment.