llm-comp-2025-subject

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.

Improvements over Standard Code

• Combined 3 datasets (~11k rows) with diverse structured output tasks
• Higher learning rate (5e-5) optimized for LoRA fine-tuning
• 2 epochs for better convergence
• Increased max_seq_len to 1024 for longer conversions
• Memory-efficient optimizer (paged_adamw_8bit)
• Gradient clipping (max_grad_norm=1.0) for training stability
• Light regularization (LoRA dropout=0.05, weight_decay=0.01)

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Method: QLoRA (4-bit)
• Max sequence length: 1024
• Epochs: 2
• Learning rate: 5e-05
• Optimizer: paged_adamw_8bit
• LoRA: r=64, alpha=128, dropout=0.05
• Gradient clipping: max_grad_norm=1.0
• CoT mask: enabled (Output: marker)

Usage

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/structured_data_with_cot_dataset_512_v5
• daichira/structured-3k-mix-sft
• daichira/structured-hard-sft-4k

Dataset License:

This dataset contains materials licensed under both the MIT License and the Creative Commons Attribution 4.0 International License (CC BY 4.0).

MIT-licensed portions are distributed under the terms of the MIT License. A copy of the license is included in the distribution.

CC BY 4.0-licensed portions are used in accordance with the Creative Commons Attribution 4.0 International License. Proper attribution to the original authors is provided, and modifications (if any) are indicated. See https://creativecommons.org/licenses/by/4.0/ for details.

Users are responsible for complying with all applicable license terms, including attribution requirements and preservation of license notices, as well as the base model’s original terms of use.