cinnamonrooo/qwen3-structeval-phase1

Qwen3-4B Structured Output LoRA (Phase 1)

This repository provides a LoRA adapter fine-tuned from
unsloth/Qwen3-4B-Instruct-2507 using QLoRA with Unsloth.

It is designed to improve the model’s ability to generate structured outputs such as:

• JSON
• YAML
• XML
• CSV
• other machine-readable formats

---

What This Repository Contains

⚠ Important

This repository contains LoRA adapter weights only.
It does not include the base model.

To use this adapter, you must load it on top of the original base model:

unsloth/Qwen3-4B-Instruct-2507

---

Training Details

Training Phase

This adapter was trained as Phase 1 using the following datasets:

• u-10bei/structured_data_with_cot_dataset_512_v2
• u-10bei/structured_data_with_cot_dataset_512_v4
• u-10bei/structured_data_with_cot_dataset_512_v5

Further training (Phase 2) may be performed later using additional datasets.

---

Training Method

• Method: QLoRA (4-bit)
• Framework: Unsloth + PEFT
• Base model: unsloth/Qwen3-4B-Instruct-2507
• Maximum sequence length: 1024
• Loss applied only to final assistant output
• Intermediate Chain-of-Thought reasoning is masked

---

Hyperparameters (Phase 1)

• LoRA rank (r): 64
• LoRA alpha: 128
• Learning rate: 1e-4
• Epochs: 1
• Batch size: 2
• Gradient accumulation: 8

---

How to Use

Example Python code to load and use this adapter:

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

base_model = "unsloth/Qwen3-4B-Instruct-2507"
adapter = "cinnamonrooo/qwen3-structeval-phase1"

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)

prompt = "Convert the following text into JSON format:\nName: John\nAge: 25"

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=200)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

License and Terms

• Training datasets: MIT License
• Base model: subject to original model license
• This adapter follows Apache 2.0 License

Users must comply with both:

1. The dataset license
2. The original base model terms

---

Notes

• This adapter is optimized for structured generation tasks
• It may not improve general conversational performance
• Designed primarily for format-following and machine-readable output accuracy

---

Future Plans

• Additional training with more datasets (Phase 2)
• Evaluation on structured output benchmarks
• Possible quantized release versions

---

If you have any questions or feedback, feel free to open an issue.