powrin/qwen3-4b-structeval-sft-merged

Model Description

This model is a merged SFT model created by combining a base model with a LoRA adapter trained for structured output generation tasks (StructEval-T style).

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Training method: Supervised Fine-Tuning (SFT) with QLoRA
• LoRA adapter source: powrin/qwen3_4b_sft_v_4ds_ep2_lr36
• Merge strategy: merge_and_unload (LoRA weights merged into base model)

This repository contains the fully merged model, ready for inference or further preference optimization (e.g. DPO).

---

Training Data

The LoRA adapter was trained using a mixture of the following datasets (only officially permitted datasets were used):

• u-10bei/structured_data_with_cot_dataset_512_v4
• u-10bei/structured_data_with_cot_dataset_512_v5
• daichira/structured-3k-mix-sft (auxiliary)
• daichira/structured-5k-mix-sft (auxiliary)

No evaluation or test data (e.g. public benchmarks) were used during training.

---

Intended Use

• Structured output generation
• JSON / schema-constrained generation
• Research on structured reasoning models
• Further fine-tuning with preference optimization (DPO)

---

Limitations

• This model is optimized for format correctness, not for free-form reasoning.
• It may underperform on open-ended or creative tasks.
• Additional tuning may be required for downstream tasks.

---

Citation

If you use this model in your research, please cite the original base model and the dataset authors accordingly.