LLM2025-SFT-LoRA

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

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) for the Matsuolab LLM 2025 competition.

Loss is applied only to the final structured output (after Output: marker), while intermediate reasoning (Chain-of-Thought) is masked.

Training Configuration

Parameter	Value
Base model	Qwen/Qwen3-4B-Instruct-2507
Method	QLoRA (4-bit, Unsloth)
Dataset	u-10bei/structured_data_with_cot_dataset_512_v2 (~3.9k)
Max sequence length	2048
Epochs	2
Learning rate	2e-6
Per-device batch size	4
Gradient accumulation	4
Effective batch size	16
Warmup ratio	0.1
Weight decay	0.05
LoRA r	64
LoRA alpha	128
LoRA dropout	0
LoRA targets	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
CoT masking	Enabled (Output: marker)
Precision	bf16

Usage

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "154teru/LLM2025-SFT-LoRA"

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 (IMPORTANT)

Training data: u-10bei/structured_data_with_cot_dataset_512_v2

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.

Framework versions

• PEFT 0.18.1