TM12/12_normal_parameter_LR5e-5

1---2base_model: Qwen/Qwen3-4B-Instruct-25073datasets:4- u-10bei/structured_data_with_cot_dataset_512_v25language:6- en7license: apache-2.08library_name: peft9pipeline_tag: text-generation10tags:11- qlora12- lora13- structured-output14---15 1612_normal_parameter_LR5e-517 18This repository provides a **LoRA adapter** fine-tuned from19**Qwen/Qwen3-4B-Instruct-2507** using **QLoRA (4-bit, Unsloth)**.20 21This repository contains **LoRA adapter weights only**.22The base model must be loaded separately.23 24## Training Objective25 26This adapter is trained to improve **structured output accuracy**27(JSON / YAML / XML / TOML / CSV).28 29Loss is applied only to the final assistant output,30while intermediate reasoning (Chain-of-Thought) is masked.31 32## Training Configuration33 34- Base model: Qwen/Qwen3-4B-Instruct-250735- Method: QLoRA (4-bit)36- Max sequence length: 51237- Epochs: 138- Learning rate: 5e-0539- LoRA: r=32, alpha=6440 41## Usage42 43```python44from transformers import AutoModelForCausalLM, AutoTokenizer45from peft import PeftModel46import torch47 48base = "Qwen/Qwen3-4B-Instruct-2507"49adapter = "TM12/12_normal_parameter_LR5e-5"50 51tokenizer = AutoTokenizer.from_pretrained(base)52model = AutoModelForCausalLM.from_pretrained(53    base,54    torch_dtype=torch.float16,55    device_map="auto",56)57model = PeftModel.from_pretrained(model, adapter)58```59 60## Sources & Terms (IMPORTANT)61 62Training data: u-10bei/structured_data_with_cot_dataset_512_v263 64Dataset License: MIT License. This dataset is used and distributed under the terms of the MIT License.65Compliance: Users must comply with the MIT license (including copyright notice) and the base model's original terms of use.66