qwen3-4b-structeval-lora

This repository provides a LoRA adapter fine-tuned from
Qwen/Qwen3-4B-Instruct-2507 for improving structured output accuracy.

⚠️ This repository contains LoRA adapter weights only.
The base model must be downloaded separately.

---

Training Objective

This LoRA adapter is trained to improve the model’s ability to generate strictly structured outputs, such as:

• JSON
• YAML
• XML
• TOML
• CSV

During training:

• Loss is applied only to the final assistant output
• Intermediate reasoning (Chain-of-Thought) is masked
• Only the content after the Output: marker is supervised

This design improves format correctness without exposing or overfitting internal reasoning traces.

---

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Training method: QLoRA (4-bit, Unsloth)
• Max sequence length: 512
• Epochs: 1
• Learning rate: 1e-6
• LoRA configuration:
  • r: 64
  • alpha: 128
• Loss type: assistant-only loss (CoT masked)

---

Usage

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

base_model = "Qwen/Qwen3-4B-Instruct-2507"
adapter_model = "hamini58/qwen3-4b-structeval-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model)

model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.float16,
    device_map="auto",
)

model = PeftModel.from_pretrained(model, adapter_model)
model.eval()

Sources & Terms (IMPORTANT)

Training dataset:
u-10bei/structured_data_with_cot_dataset_512_v2

Dataset License:
MIT License

Compliance:
Users must comply with:

• The MIT License of the training dataset (including copyright notice)
• The original license and terms of use of the base model
  (Qwen/Qwen3-4B-Instruct-2507, Apache License 2.0)

This repository distributes only LoRA adapter weights and does not redistribute the base model.