Qwen3.5-27B WebArena NoSkill Agentic-ESOpt

This is the final full-parameter model produced by the NoSkill + Agentic-ESOpt WebArena run.

  • Base model: Qwen/Qwen3.5-27B
  • Training: Agentic-ESOpt, population 8, 70 generations
  • Update: full parameters, alpha 2.5e-4, z-score reward normalization
  • Exploration sigma: 1.5e-3
  • WebArena evaluation: 36.16% mean over three runs

The repository contains standard Transformers safetensors weights and tokenizer files.

Downloads last month
151
Safetensors
Model size
27B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for zz1358m/Qwen3.5-27B-WebArena-Agentic-ESOpt

Base model

Qwen/Qwen3.5-27B
Finetuned
(296)
this model

Collection including zz1358m/Qwen3.5-27B-WebArena-Agentic-ESOpt