Qwen-4B-DB-AlfWorld-v16

Qwen-4B-DB-AlfWorld-v16

This repository provides a merged model fine-tuned from Qwen/Qwen3-4B-Instruct-2507 on datasets u-10bei/sft_alfworld_trajectory_dataset_v5, dbbench_sft_dataset_react_v2 and dbbench_sft_dataset_react_v3.

All LoRA adapter weights have been merged into the base model, and the resulting merged model is saved here as a standalone model. No external adapter loading is required.

Dataset Notes (IMPORTANT)

ALFWorld datasets

For the datasets:

• u-10bei/sft_alfworld_trajectory_dataset_v5
• u-10bei/sft_alfworld_trajectory_dataset_v4

the following preprocessing steps were applied:

1. Only samples that include admissible actions for each dialogue turn were extracted. This ensures high-quality supervision aligned with the agent’s available action space.

2. Inserted the prefix below at the beginning of the first assistant message: Task Type: AGENT This explicitly marks the trajectory as an agent-based task.

Additionally, for sft_alfworld_trajectory_dataset_v5, only samples with input length ≤ 2048 tokens were used during training to ensure training stability and consistency with the maximum sequence length.

DBBench dataset

For u-10bei/dbbench_sft_dataset_react_v3, the following preprocessing was applied:

• Inserted the prefix: Task Type: DATABASE at the start of the assistant’s initial turn. This makes the task type explicit and improves instruction consistency.

Training Objective

This model is trained to improve multi-turn agent task performance on ALFWorld (household tasks) and DBBench (database operations).

Loss is applied to all assistant turns in multi-turn trajectories, enabling the model to learn observation interpretation, step-by-step reasoning, action execution, tool use, and recovery from errors.

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Method: LoRA (merged into final weights)
• Max sequence length: 2048
• Learning rate: 1e-06
• LoRA parameters used during training: r=32, alpha=64

Usage (Agent-style Inference Example)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Umiharu/Qwen-4B-DB-AlfWorld-v16"


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

prompt = "You are a household task-solving agent. Respond 'OK' if you are ready."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=64,
    temperature=0.2,
    do_sample=False,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Sources & Terms (IMPORTANT)

Training data: u-10bei/sft_alfworld_trajectory_dataset_v5, dbbench_sft_dataset_react_v2 and dbbench_sft_dataset_react_v3

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.