Kaito-F/qwen3-4b-agentbench-opd-adapter-v2-sample

Qwen3-4B AgentBench LoRA (SFT + OPD)

This LoRA adapter was trained in two stages:

1. SFT (Supervised Fine-Tuning) on agent trajectory data
2. OPD (On-Policy Distillation) from Qwen/Qwen3-30B-A3B-Instruct-2507

Training Configuration

Stage 1: SFT

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Dataset: ALFWorld trajectories

Stage 2: OPD

• Teacher model: Qwen/Qwen3-30B-A3B-Instruct-2507
• Dataset: ALFWorld + DBBench (combined)
• Steps: 100
• Learning rate: 1e-05

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "Kaito-F/qwen3-4b-agentbench-opd-adapter-v2-sample"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)