qwen3-4b-agentbench-dbalf-lora

This repository provides a LoRA adapter fine-tuned from Qwen/Qwen3-4B-Instruct-2507 using LoRA + Unsloth for AgentBench-style multi-turn agent trajectories.

This repository contains LoRA adapter weights only. The base model must be loaded separately.

Training Objective

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

• DBBench (database operation / SQL generation trajectories)
• ALFWorld (household task trajectories)

Loss is applied to all assistant turns in the trajectory, enabling the model to learn:

• environment observation
• action selection
• tool use / operation formatting
• recovery from intermediate errors

Training Data

• DBBench dataset: u-10bei/dbbench_sft_dataset_react_v4
• ALFWorld dataset: u-10bei/sft_alfworld_trajectory_dataset_v5
• Mixing ratio (pre-merge target): DB:ALF = 1:0

DB Oversampling (category-aware)

Enabled: False

DB category weights used during training-data preparation:

• counting: 1
• comparison: 1
• ranking: 1
• select: 1
• insert: 1
• update: 1
• other: 1

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Method: LoRA (full precision base)
• Max sequence length: 2048
• Epochs: 1.2
• Learning rate: 3e-06
• LoRA: r=32, alpha=64, dropout=0.0
• Per-device train batch size: 2
• Gradient accumulation: 4

Usage

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "AF0815/agentbench"

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

Notes

• This repository is intended for adapter-only distribution.
• Please ensure compliance with the base model license/terms in addition to this repository's license.
• If you publish evaluation results, it is recommended to report:
  • AgentBench task split / seeds
  • DBBench / ALFWorld mix ratio
  • DB oversampling settings
  • decoding settings

Sources & Terms (IMPORTANT)

Training data:

• u-10bei/dbbench_sft_dataset_react_v4
• u-10bei/sft_alfworld_trajectory_dataset_v5

Dataset license / terms:

• Please follow the original license and terms of each dataset repository.
• This adapter repository license (apache-2.0) applies to the adapter files in this repository.