legoskier/qwen3-4b-agent-trajectory-lora_2

qwen3-4b-agent-trajectory-lora

This repository provides a LoRA adapter fine-tuned from
Qwen/Qwen3-4B-Instruct-2507 using LoRA + Unsloth.

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

Training Objective

Trained to improve multi-turn agent task performance on
ALFWorld (household tasks) and DBBench (database operations).
Loss is applied to all assistant turns, enabling the model to learn
environment observation, action selection, tool use, and error recovery.

Training Configuration

Item	Value
Base model	Qwen/Qwen3-4B-Instruct-2507
Method	LoRA (full precision base, bf16)
Datasets	ALFWorld v5 + DBBench v4
Max sequence length	8192
Epochs	2
Learning rate	5e-06
LoRA r / alpha	128 / 256

Usage

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

base    = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "legoskier/qwen3-4b-agent-trajectory-lora_2"

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

Sources & Terms

Training datasets:

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

This repository does not redistribute the datasets.
Users must comply with the respective dataset licenses and the base model's terms of use.