azuki-digital/qwen3-4b-struct-lora-v11-adapter

qwen3-4b-struct-lora-v11

This repository provides a LoRA adapter fine-tuned on top of:

azuki-digital/qwen3-4b-struct-lora-v4-merged

⚠️ 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

Item	Value
Base model	azuki-digital/qwen3-4b-struct-lora-v4-merged
Method	LoRA SFT (no quantization, bf16)
Max sequence length	4096
Epochs	2
Learning rate	1e-5
Warmup ratio	0.05
Weight decay	0.05
LoRA r	32
LoRA alpha	64
LoRA dropout	0.05
Target modules	q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
Mask CoT	Yes (after_marker)
Dataset	daichira/structured-3k-mix-sft

---

Usage

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

base = "azuki-digital/qwen3-4b-struct-lora-v4-merged"
adapter = "azuki-digital/qwen3-4b-struct-lora-v11"

tokenizer = AutoTokenizer.from_pretrained(base)

model = AutoModelForCausalLM.from_pretrained(
    base,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

model = PeftModel.from_pretrained(model, adapter)

Sources & License (IMPORTANT)

Training Data • daichira/structured-3k-mix-sft • License: CC-BY-4.0

This dataset is used under the terms of CC-BY-4.0.

Compliance

Users must comply with: 1. Dataset attribution requirements (CC-BY-4.0) 2. Base model license (Apache-2.0) 3. This repository license (Apache-2.0)