You can deploy this model and start earning money today!

harutoshi/
qwen3-4b-sft-structured-hard-4k

Name: qwen3-4b-sft-structured-hard-4k
Availability: InStock

daichira/structured-hard-sft-4k

Qwen/Qwen3-4B-Instruct-2507

adapter

apache-2.0

Views

No views yet

Qwen3-4b-SFT-Structured-Hard-4k

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

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) even for complex and long (up to 4k tokens) tasks.

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

Training Configuration

Base model: Qwen/Qwen3-4B-Instruct-2507
Method: QLoRA (4-bit)
Max sequence length: 4096
Epochs: 1
Learning rate: 2e-04
LoRA: r=64, alpha=128
Dataset: daichira/structured-hard-sft-4k

Usage

python

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "harutoshi/qwen3-4b-sft-structured-hard-4k"

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: daichira/structured-hard-sft-4k 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.