CoolFace
Modelpublic

kuririrn/qwen3-4b-structured-output-lora-tuned_param_v2-without_cot

sourceHugging Faceapache-2.0updated 8mo agoView on Hugging Face
0likes12downloads
Model Card

qwen3-4b-structured-output-lora-tunedparamv2-without_cot

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).

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: 512
  • —Epochs: 3
  • —Learning rate: 1e-06
  • —LoRA: r=64, alpha=128
  • —Task-weighted sampling: enabled (see below)

Data preprocessing / curation

The training data is derived from u-10bei/structured_data_with_cot_dataset_512_v2 with the following preprocessing:

  • —Keep only samples whose final message is a non-empty assistant turn.
  • —Identify the final assistant message and locate the last occurrence of one of the output markers: Output:, OUTPUT:, Final:, Answer:, Result:, Response:.
  • —If such a marker is found, the assistant content is trimmed so that only the substring after the marker is retained as the training target.
  • —When the retained segment is enclosed in a Markdown-style code fence (e.g. ``json, `yaml, `xml, ``toml), the outer code fence is removed, and only the raw structured content is kept.
  • —If no output marker is present, the original assistant content is preserved.

This preprocessing is applied uniformly to both generation and conversion tasks and is performed only at training time. No post-processing or output modification is applied during inference.

Usage

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "your_id/your-repo"

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: u-10bei/structureddatawithcotdataset512v2

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.