takayosh/mix2

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

This adapter is trained on mixed multi-turn agent trajectories from:

• ALFWorld datasets
• DBBench datasets

The datasets are mixed with:

• ALF ratio: 0.9
• DB ratio: 0.1

Data preprocessing includes:

• Marker normalization (THOUGHT / ACTION / ANSWER)
• Removal of unnecessary task-completion lines
• Optional ACTION-only training
• Optional CoT masking
• Assistant-only loss

---

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Method: LoRA (full precision base)
• Max sequence length: 2048
• Epochs: 2
• Learning rate: 1e-05
• LoRA: r=128, alpha=256

Usage

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "takayosh/mix2"

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.