erv1n/e2l-alfworld-qwen3-8b-coldstart

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The erv1n/e2l-alfworld-qwen3-8b-coldstart model is an 8 billion parameter Qwen3-based checkpoint, specifically designed as a cold-start initialization for policy training in the ALFWorld environment. Developed by erv1n, it leverages online mistake-driven experience learning (next-observation SFT gated by a semantic judge) on ALFWorld's AlfredTWEnv. This model is primarily intended to provide an initial state for Stage-2 policy GRPO training, offering a specialized foundation for reinforcement learning agents in text-based game environments.

Loading preview...

erv1n/e2l-alfworld-qwen3-8b-coldstart Overview

This model is an 8 billion parameter checkpoint based on the Qwen3 architecture, developed by erv1n. It serves as a "cold-start" initialization for more advanced policy training, specifically within the ALFWorld environment using AlfredTWEnv.

Key Capabilities & Training

  • Specialized Initialization: Designed to provide an initial state for Stage-2 policy GRPO (Gradient-based Reinforcement Policy Optimization) training.
  • Online Mistake-Driven Learning: The model's training incorporates online mistake-driven experience learning, where next-observation Supervised Fine-Tuning (SFT) is gated by a semantic judge.
  • ALFWorld Optimization: Fine-tuned specifically for the ALFWorld AlfredTWEnv environment, indicating a focus on text-based game and interactive agent tasks.
  • Hugging Face Compatibility: Can be loaded as a standard Hugging Face checkpoint using AutoModelForCausalLM.from_pretrained.

Intended Use Cases

  • Reinforcement Learning Research: Ideal for researchers and developers working on reinforcement learning agents in text-based environments like ALFWorld.
  • Policy Training Initialization: Primarily used as a foundational checkpoint to accelerate and improve the starting point for subsequent policy training phases, particularly with systems like verl-agent.

Finetunes
1 models