erv1n/e2l-alfworld-qwen25-7b-coldstart
The erv1n/e2l-alfworld-qwen25-7b-coldstart model is a 7.6 billion parameter checkpoint based on the Qwen/Qwen2.5-7B-Instruct architecture. It serves as a stage-1 cold-start initialization for online mistake-driven experience learning on ALFWorld's AlfredTWEnv. This model is specifically designed to be the initial checkpoint for Stage-2 policy GRPO training within the verl-agent framework, making it suitable for research in reinforcement learning and interactive environments.
Loading preview...
Model Overview
The erv1n/e2l-alfworld-qwen25-7b-coldstart is a specialized 7.6 billion parameter language model, derived from the Qwen/Qwen2.5-7B-Instruct architecture. It represents a stage-1 "cold-start" checkpoint, specifically developed for online mistake-driven experience learning within the ALFWorld AlfredTWEnv environment.
Key Capabilities and Purpose
- Initialization for Reinforcement Learning: This model's primary function is to serve as an initial state for more advanced training stages. It is designed to be loaded as a regular Hugging Face checkpoint (
AutoModelForCausalLM.from_pretrained) to kickstart subsequent training. - Online Mistake-Driven Learning: The checkpoint is a result of full-parameter online mistake-driven experience learning, where next-observation Supervised Fine-Tuning (SFT) is gated by a semantic judge. This process aims to improve the model's performance in interactive, text-based environments.
- ALFWorld Environment Focus: The training and intended use are specifically tailored for the ALFWorld
AlfredTWEnv, indicating its specialization in text-based game environments and interactive decision-making.
Intended Use Case
This model is specifically intended for use as the initialization point for Stage-2 policy GRPO (Generalized Policy Replay Optimization) training within the verl-agent framework. Researchers and developers working on reinforcement learning agents for interactive text environments, particularly those utilizing the verl-agent system, will find this checkpoint valuable for setting up their training pipelines. The controller_state.json file included with the model contains experiment-controller provenance metadata and can typically be ignored for standard inference or training setups.