qwen3-4b-sft-lora-v5-2-20260208-1513

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

LoRA adapter weights only - base model must be loaded separately.

Training Objective

Structured output accuracy (JSON / YAML / XML / TOML / CSV). Loss applied only to final assistant output (CoT masked).

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Method: QLoRA (4-bit)
• Max sequence length: 512
• Epochs: 1
• Learning rate: 5e-06
• LoRA: r=64, alpha=128
• Weight decay: 0.05
• Warmup ratio: 0.1
• Effective batch size: 16

Usage (adapter_merge mode)

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "yusei926/qwen3-4b-sft-lora-v5-2-20260208-1513"

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

Training data: u-10bei/structured_data_with_cot_dataset_512_v5 Dataset License: MIT License.