CoolFace
Modelpublic

Sarim-Hash/browseragent-rft-web-coevo-executors

sourceHugging Faceupdated 2mo agoView on Hugging Face
0likes
Model Card

BrowserAgent-RFT web co-evolution executors

LoRA adapters (r=16, α=32) over TIGER-Lab/BrowserAgent-RFT (Qwen2.5-7B), produced by an Agent0-style web co-evolution loop: a curriculum LoRA proposes <PAGE>/<GOAL> tasks and this executor LoRA solves them multi-turn inside a frozen WebWorld world model, trained with GRPO using a label-free self-consistency reward.

Adapters

subfolderiteration
exec_v1co-evolution iter 1
exec_v2co-evolution iter 2
exec_v3co-evolution iter 3 (final)

Result (held-out hard-20, 5-seed, strict final_click)

base 0.640  →  exec_v1 0.740  →  exec_v2 0.790  →  exec_v3 0.880   (+0.24)

Usage

python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("TIGER-Lab/BrowserAgent-RFT")
tok  = AutoTokenizer.from_pretrained("TIGER-Lab/BrowserAgent-RFT")
model = PeftModel.from_pretrained(base, "Sarim-Hash/browseragent-rft-web-coevo-executors", subfolder="exec_v3")

These are adapters only — the base model TIGER-Lab/BrowserAgent-RFT is required.