nakotsuko13/qwen3-4b-nako13-agentbench-lora

Agent SFT Model (ALFWorld & DBBench)

This repository provides a LoRA adapter fine-tuned from Qwen/Qwen3-4B-Instruct-2507 using LoRA + Unsloth.

Training Objective

This adapter is trained to improve multi-turn agent task performance on:

1. ALFWorld: Household interaction tasks.
2. DBBench: Complex SQL database operations.

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Max sequence length: 2048
• Epochs: 2
• Learning rate: 2e-06
• LoRA: r=64, alpha=128

Training Data

• nakotsuko13/distilled_dataset_v2_clean

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = "Qwen/Qwen3-4B-Instruct-2507"
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(model, "your_id/your-repo")