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 & Merge Configuration

[Merge Settings (MergeKit)]

Method: TIES
Base model for merge: Qwen/Qwen3-4B-Instruct-2507
Models & Parameters:
maru-miya/lora_agentbench_qwen3_4b_d20_t1 (weight: 1.0, density: 0.9)
maru-miya/lora_agentbench_qwen3_4b_d21_t9_db (weight: 0.5, density: 0.3)

[Original LoRA Adapter 1: d20_t1]

Base model: Qwen/Qwen3-4B-Instruct-2507
Method: LoRA (full precision base)
Max sequence length: 2048
Epochs: 2
Learning rate: 2e-06
LoRA: r=16, alpha=32

[Original LoRA Adapter 2: d21_t9_db]

Base model: Qwen/Qwen3-4B-Instruct-2507
Method: LoRA (full precision base)
Max sequence length: 2048
Epochs: 2
Learning rate: 2e-05
LoRA: r=16, alpha=32

Usage

python

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "maru-miya/merged_agentbench_qwen3_4b_dare_ties_v2"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(
    base,
    torch_dtype=torch.float16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter)
Sources & Terms (IMPORTANT)
Training data:

u-10bei/sft_alfworld_trajectory_dataset_v5

u-10bei/dbbench_sft_dataset_react

This repository does NOT redistribute the dataset.
Users must comply with the dataset license and base model terms.