chanyoungkim/ppo-qwen2.5-1.5b-alfworld-step130

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 7, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The chanyoungkim/ppo-qwen2.5-1.5b-alfworld-step130 is a 1.54 billion parameter Qwen2.5-Instruct model fine-tuned using Proximal Policy Optimization (PPO) for the ALFWorld environment. This specific checkpoint, at optimizer step 130, demonstrates a 58.33% success rate on ALFWorld valid_seen tasks, significantly outperforming the base model's 2.86%. It is optimized for sequential decision-making and interactive agent tasks within text-based environments.

Loading preview...

Overview

This model is a 1.54 billion parameter Qwen2.5-Instruct backbone, fine-tuned using Proximal Policy Optimization (PPO) for the ALFWorld environment. Developed by chanyoungkim, it leverages the langfengQ/verl-agent framework for reinforcement learning. This particular release represents the model at optimizer step 130, showcasing its performance after targeted training for interactive tasks.

Key Capabilities

  • Enhanced Task Success: Achieves a 58.33% success rate on 128 ALFWorld valid_seen tasks, a substantial improvement over the base Qwen2.5-1.5B-Instruct model's 2.86%.
  • Efficient Action Generation: Demonstrates a high action format rate of 96.5%, ensuring parseable outputs for environmental interaction.
  • Optimized for Interactive Environments: Specifically trained to navigate and complete tasks within the ALFWorld text-based environment, indicating proficiency in sequential decision-making.
  • PPO Fine-tuning: Utilizes PPO with specific hyperparameters, including KL loss and an invalid action penalty, to guide policy learning.

Good for

  • ALFWorld Benchmarking: Ideal for researchers and developers working on agent-based models within the ALFWorld environment, offering a strong baseline and checkpoint for further experimentation.
  • Reinforcement Learning Research: Provides a practical example of PPO application for fine-tuning large language models in interactive settings.
  • Developing Interactive Agents: Suitable for use cases requiring an LLM to act as an agent in text-based games or simulated environments where sequential actions are critical.

Finetunes
1 models