qwen3-4b-structured-output-lora-continued-v5-ver3

qwen3-4b-structured-output-lora-continued-v5-ver3

A LoRA adapter specialized for structured output generation (JSON / YAML / XML / TOML / CSV) in long-input settings.

This repository provides a LoRA adapter fine-tuned from Qwen/Qwen3-4B-Instruct-2507 using QLoRA (4-bit, Unsloth).

This adapter was trained in five stages using LoRA continuation: structured_data_with_cot_dataset_512_v2 (initial fine-tuning) → structured_data_with_cot_dataset_512_v5 (continued fine-tuning) → structured-5k-mix-sft (continued fine-tuning) → mixed training (continued fine-tuning) using structured-5k-mix-sft as main data and structured-hard-sft-4k as auxiliary data.→ structured_data_with_cot_dataset_512_v5 (continued fine-tuning(retraining))

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. When output markers are present, intermediate reasoning (Chain-of-Thought) is masked.

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Method: QLoRA (4-bit)
• Max sequence length: 1024
• Epochs: 2
• Learning rate: 5e-07
• LoRA: r=64, alpha=128

Usage

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

base = "Qwen/Qwen3-4B-Instruct-2507"
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 (stage 1): u-10bei/structured_data_with_cot_dataset_512_v2
• Training data (stage 2): u-10bei/structured_data_with_cot_dataset_512_v5
• Training data (stage 3): daichira/structured-5k-mix-sft
• Training data (stage 4): mixed dataset
  • Main: daichira/structured-5k-mix-sft
  • Auxiliary: daichira/structured-hard-sft-4k
• Training data (stage 5): u-10bei/structured_data_with_cot_dataset_512_v5(Retraining)

Dataset License & Attribution

The following licenses apply to the training datasets used in this repository:

• u-10bei structured_data_with_cot_dataset family (e.g., structured_data_with_cot_dataset_512_v2, v4, v5, etc.): Licensed under the MIT License. Use of these datasets is permitted under the terms of the MIT License, provided that the copyright notice is retained.

• daichira structured_ SFT datasets* (structured-5k-mix-sft, structured-hard-sft-4k): Licensed under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license. This license permits use, modification, and redistribution, including for commercial purposes, provided that appropriate attribution is given. Attribution: please credit the original dataset authors (u-10bei, daichira) and link to the dataset pages when redistributing or publishing derived works.

Users must comply with:

1. The license terms of each individual dataset listed above.
2. The original license and terms of use of the base model (Qwen/Qwen3-4B-Instruct-2507).

This repository distributes LoRA adapter weights only and does not redistribute the original datasets.