mkd-hika/qwen2.5-7b-instruct-alpaca-ft

Fine-tuned from Qwen/Qwen2.5-7B-Instruct on the alpaca_guanaco-llama2-1k (50877 rows) dataset.

Intended use

General-purpose instruction-following assistant fine-tuned for enterprise AI applications at MKD Co., Ltd. Suitable for Korean and English conversational tasks, document Q&A, and structured instruction completion.

Training details

• Method: QLoRA (LoRA rank 32, NF4 base weights, bfloat16 compute)
• Hardware: NVIDIA GeForce RTX 5090 31GB
• Duration: pending
• Dataset: alpaca_guanaco-llama2-1k (50877 rows)
• Base model: Qwen/Qwen2.5-7B-Instruct
• Final val loss: 0.2851
• Trained by: Hika — MKD Co., Ltd.

Evaluation

Full report: eval_results/run_2026-06-09_0540/summary.md

Eval	Base	This model	Δ
MMLU	pending	pending	pending
MT-Bench	pending	pending	pending

Bias, risks, limitations

• Fine-tuned on English-centric instruction data (Alpaca/Guanaco); performance on Korean-specific tasks may be limited without Korean dataset fine-tuning.
• Not evaluated for safety or harmful content — do not deploy in production without additional alignment and safety testing.
• Inherits all limitations of the base Qwen2.5-7B-Instruct model.

License

Inherits license of Qwen/Qwen2.5-7B-Instruct.

Citation

If you use this model in research or products, please cite: ``` @misc{mkd-qwen2.5-7b-alpaca-ft, author = {Hika, MKD Co., Ltd.}, title = {Qwen2.5-7B-Instruct fine-tuned on Alpaca/Guanaco}, year = {2026}, publisher = {HuggingFace}, url = {https://huggingface.co/mkd-hika/qwen2.5-7b-instruct-alpaca-ft} } ```