qwen3-4b-agent-trajectory-lora-v18

qwen3-4b-agent-trajectory-lora-v18

This repository provides a merged full model derived from:

• Qwen/Qwen3-4B-Instruct-2507
• yokobo-ai/qwen3-4b-agent-trajectory-lora-v17
  The v18 model is obtained by further fine-tuning v17 on additional agent trajectory data. This repository contains a fully merged model (base + LoRA merged). No separate adapter loading is required.

---

Model Lineage

Qwen3-4B-Instruct-2507
→ LoRA fine-tuning (v17)
→ Additional fine-tuning (v18, merged full model)

Training Objective

The model is optimized for multi-turn agent tasks including:

• ALFWorld (household interaction tasks)
• DBBench (database reasoning & SQL tasks) The goal of v18 is to improve database task performance while preserving strong ALFWorld capabilities. Loss is applied to all assistant turns in trajectory data, enabling learning of:
• environment observation
• action selection
• tool use
• recovery from errors

---

Training Configuration

• Starting checkpoint: yokobo-ai/qwen3-4b-agent-trajectory-lora-v17
• Original base model: Qwen/Qwen3-4B-Instruct-2507
• Method: LoRA fine-tuning + merge
• Max sequence length: 2048
• Epochs: 1
• Learning rate: 1e-06
• LoRA: r=64, alpha=128

Usage

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

base = "yokobo-ai/qwen3-4b-agent-trajectory-lora-v17"
adapter = "your_id/your-repo"

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_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.