moushi21/agent-bench-merged12

Performance

The model shows balanced performance across agentic tasks:

• ALFWorld: 0.60 (Pass@1)
• DBBench: 0.5353

Qwen3-4B Agent TIES Merge v1.2

This repository provides a merged model optimized for agentic tasks, created by merging specialized LoRA adapters into the base model Qwen/Qwen3-4B-Instruct-2507.

The merge was performed using the TIES-Merging method via Mergekit, combining expertise in ALFWorld trajectories and DBBench (SQL) tasks.

This repository contains full model weights, ready for inference without needing to load separate adapters.

Merge Details

This model was merged using the TIES method.

Merge Method

• Method: ties
• Base Model: Qwen/Qwen3-4B-Instruct-2507
• Dtype: bfloat16

Configuration

The following mergekit-yaml configuration was used to produce this model:

merge_method: ties
base_model: Qwen/Qwen3-4B-Instruct-2507
dtype: bfloat16
models:
  - model: moushi21/agent-bench-alfworld-merged3
    parameters:
      weight: 1.0    # ALF Main
      density: 0.3   # 70% Noise Cut
  - model: moushi21/agent-bench-dbbench-merged4
    parameters:
      weight: 0.3    # DB
      density: 0.3   # 70% Noise Cut

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "moushi21/agent-bench-merged12"

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

Sources & Terms (IMPORTANT)

Training Data (Source Model Lineage)

The source models utilized the following datasets for fine-tuning:

• ALFWorld: u-10bei/sft_alfworld_trajectory_dataset (v1 to v5)
• DBBench: u-10bei/dbbench_sft_dataset_react (v1 to v4)

License & Compliance

• Dataset License: MIT License. These datasets are 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 (Qwen3-4B-Instruct-2507).