Qwen3-4B-SFT

Qwen3-4B-SFT (STEP100 Checkpoint)

This repository provides the LoRA adapter checkpoint at training step 100
fine-tuned from:

Qwen/Qwen3-4B-Instruct-2507

This is the best-performing checkpoint selected from training.

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

---

Training Overview

This adapter was trained using:

• Supervised Fine-Tuning (SFT)
• QLoRA (4-bit, Unsloth)
• Output-only loss (Chain-of-Thought masked)
• Task-aware upsampling

Checkpoint:

• Selected step: 100
• Training max steps: 200
• Best validation performance observed at step 100

---

Training Objective

The goal is to improve structured output accuracy across:

• JSON
• YAML
• XML
• TOML
• CSV

Output-only Supervision

Loss is applied only to the final assistant output.

Intermediate reasoning (Chain-of-Thought) is masked.

Learning begins after output markers such as: Output: OUTPUT: Final: Answer: Result: Response:

Configuration:

• SFT_MASK_COT = 1
• SFT_OUTPUT_LEARN_MODE = "after_marker"

---

Task-aware Upsampling

Upsampling was enabled to rebalance structured transformation tasks.

Applied multipliers:

Configuration:

• SFT_USE_UPSAMPLING = 1

This improves XML-related transformation robustness while maintaining multi-format generalization.

---

Training Configuration

Base Model

Qwen/Qwen3-4B-Instruct-2507

Dataset

u-10bei/structured_data_with_cot_dataset_512_v2

Method

QLoRA (4-bit) via Unsloth

Hyperparameters

---

LoRA Configuration

---

Repository Contents

This repository contains only:

• adapter_config.json
• adapter_model.safetensors
• README.md

This is the STEP100 checkpoint only (final selected model).

---

Usage

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

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "Gen-oze/Qwen3-4B-SFT"

tokenizer = AutoTokenizer.from_pretrained(base)

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

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