moushi21/agent-bench-dbbench-merged4

Qwen3-4B-Agent-DBBench-Specialist

This repository provides a merged full-parameter model (bfloat16) fine-tuned from Qwen/Qwen3-4B-Instruct-2507.

Instead of a standalone LoRA adapter, this model has been created by merging LoRA weights back into the base model using Unsloth's merge_and_unload method. This ensures high-speed inference and easy deployment.

Training Objective

This model is specialized for DBBench trajectory tasks, trained to handle multi-turn environment observations and action selections.

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Format: Merged Full Weights (bfloat16)
• Method: LoRA fine-tuning (Merged via Unsloth merge_and_unload)
• Max sequence length: 4096
• Steps: 500
• Learning rate: 5e-07
• LoRA Parameters during training: r=64, alpha=128
• Platform: Trained with Unsloth

Usage

Since this is a merged model, you can load it directly like any other Qwen3 model:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "moushi21/agent-bench-dbbench-merged4"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

Sources & Terms (IMPORTANT)

Training data:

• u-10bei/dbbench_sft_dataset_react
• u-10bei/dbbench_sft_dataset_react_v2
• u-10bei/dbbench_sft_dataset_react_v3
• u-10bei/dbbench_sft_dataset_react_v4

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.