This is an improved version of the standard SFT training code. Key changes from the baseline are based on token length analysis of the training dataset.

Changes from Baseline

Parameter	Baseline	v1	Rationale
MAX_SEQ_LEN	512	1024	Token analysis: P99=640-961. 512 truncates data
Epochs	1	3	Small dataset (~3.6k rows) benefits from more passes
Learning Rate	1e-6	2e-05	Higher LR is effective for LoRA fine-tuning
Batch Size	2	4	L4/A100 has sufficient VRAM
Grad Accum	8	4	Reduced to maintain effective BS=16

Training Objective

This adapter is trained to improve structured output accuracy (JSON / YAML / XML / TOML / CSV) for the StructEval-T benchmark.

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, Unsloth)
Max sequence length: 1024
Epochs: 3
Learning rate: 2e-05
Batch size: 4 (effective: 16)
Gradient accumulation: 4
LoRA: r=64, alpha=128
CoT masking: enabled (loss on final output only)

Usage

python

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "your_id/qwen3-4b-structured-output-lora-v1"

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_v2

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.

kmd2525/qwen3-4b-structured-output-lora-v1

qwen3-4b-structured-output-lora-v1

Version: v1 — Hyperparameter Improvement

Changes from Baseline

Training Objective

Training Configuration

Usage

Sources & Terms (IMPORTANT)

kmd2525/
qwen3-4b-structured-output-lora-v1