This repository provides a LoRA adapter obstained by continued fine-tuning of an existing structured-output LoRA originally trained on unsloth/Qwen3-4B-Instruct-2507.

This adapter is not trained from scratch. It is a continuation of:

mandalgo/qwen3-4b-structured-output-lora_9

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

Training Objective

This adapter is trained to improve structured output accuracy (JSON / YAML / XML / TOML / CSV).

Loss is applied only to the final assistant output, while intermediate reasoning (Chain-of-Thought) is masked.

Training Configuration

Base model: unsloth/Qwen3-4B-Instruct-2507
Parent adapter: mandalgo/qwen3-4b-structured-output-lora_9
Method: QLoRA (4-bit)
Max sequence length: 512
Epochs: 1 (continued fine-tuning)
Learning rate: 2e-07
LoRA: r=64, alpha=128

Usage

python

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

base = "unsloth/Qwen3-4B-Instruct-2507"
adapter = "mandalgo/qwen3-4b-structured-output-lora_21"

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/structured_data_with_cot_dataset_512_v5 (MIT License)
daichira/structured-hard-sft-4k (CC-BY-4.0, synthetic)

The daichira dataset is a purely synthetic dataset licensed under Creative Commons Attribution 4.0 (CC-BY-4.0).

When using or redistributing this adapter, users must provide appropriate attribution to the dataset author in accordance with the CC-BY-4.0 license.

This adapter is a continued fine-tuning of a prior LoRA adapter and must also comply with:

The base model license (Apache-2.0)
The original adapter license
All dataset licenses listed above

Attribution

Dataset "structured-hard-sft-4k" by daichira, licensed under CC-BY-4.0.