SingleBicycle/ccpo-alfworld-qwen2.5-7b-global-fa

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SingleBicycle/ccpo-alfworld-qwen2.5-7b-global-fa is an interim checkpoint of a 7.6 billion parameter agent policy based on Qwen2.5-7B-Instruct, fine-tuned for credit assignment in multi-turn agent reinforcement learning using Context-Conditioned Policy Optimization (CCPO) on the ALFWorld text-based embodied household benchmark. This specific arm utilizes a global gate and successor-node target, achieving a held-out success rate of 77.5% on unseen ALFWorld tasks. It is designed for research into advanced RL techniques for complex, multi-step environments.

Loading preview...

Overview

This model, SingleBicycle/ccpo-alfworld-qwen2.5-7b-global-fa, is an interim checkpoint of an agent policy trained with Context-Conditioned Policy Optimization (CCPO) on the ALFWorld text-based embodied household benchmark. Built upon the Qwen2.5-7B-Instruct base model, it focuses on improving credit assignment in multi-turn agent reinforcement learning (RL).

CCPO enhances traditional RL methods by replacing the step term of the advantage function with a context-conditioned, uncertainty-shrunk leave-one-out baseline. This allows for more precise credit assignment by isolating the estimator's contribution.

Key Capabilities

  • Advanced RL Agent: Implements CCPO for sophisticated credit assignment in multi-turn decision-making.
  • ALFWorld Proficiency: Achieves a 77.5% held-out success rate on unseen ALFWorld tasks (converged window mean), demonstrating strong performance in complex text-based environments.
  • Qwen2.5-7B-Instruct Base: Leverages the capabilities of a 7.6 billion parameter instruction-tuned language model.
  • Research Focus: Designed for exploring and validating novel RL estimation techniques, specifically the global gate and successor-node target approach.

Good for

  • Reinforcement Learning Research: Ideal for researchers studying credit assignment, policy optimization, and agent behavior in complex, sequential environments.
  • ALFWorld Benchmark: Suitable for evaluating and comparing advanced agent policies on the ALFWorld benchmark.
  • Understanding CCPO: Provides a practical implementation for analyzing the effects of context-conditioned baselines in RL.

Note: This is an interim checkpoint (step 111 of 150). Final weights and settled numbers will be updated upon completion of the training run. The results presented are from a single seed and are not directly comparable to published ALFWorld results using smaller base models (e.g., Qwen2.5-1.5B-Instruct).

Finetunes
1 models