qwen3-4b-agent-v3-lora-v2

This repository provides a merged model fine-tuned from SELEE/qwen3-4b-agent-v3 using LoRA + Unsloth.

Because the LoRA weights have been merged, this repository contains the full model weights. The model can be loaded directly without needing to load the base model and adapter separately.

Training Objective

The primary objective of this fine-tuning phase is to further improve performance on DBBench (database operations). To avoid catastrophic forgetting of ALFWorld, a portion of the ALFWorld dataset is included as replay during training.

Loss is applied to all assistant turns in the multi-turn trajectory, enabling the model to learn environment observation, action selection, tool use, and recovery from errors.

Training Configuration

• Base model: SELEE/qwen3-4b-agent-v3
• Method: LoRA (merged into full precision base)
• Max sequence length: 4096
• Epochs: 2
• Learning rate: 2e-06
• LoRA: r=128, alpha=128

Usage

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

base = "SELEE/qwen3-4b-agent-v3-lora-v2"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(
    base,
    torch_dtype=torch.float16,
    device_map="auto",
)

Sources & Terms (IMPORTANT)

Training data: u-10bei/dbbench_sft_dataset_react_v4, u-10bei/sft_alfworld_trajectory_dataset_v5

Dataset License: MIT License. This dataset is used and distributed under the terms of the MIT License. Compliance: Users must comply with the MIT license (including copyright notice) and the base model's original terms of use.