Model Card for LDCC-Instruct-Llama-2-ko-13B-v1.4

LDCC-Instruct-Llama-2-ko-13B-v1.4 is a continuation in a series of language models designed to serve as efficient assistants. This fifth iteration is an enhanced version of its predecessor, LDCC/LDCC-Instruct-Llama-2-ko-13B-v1.0. We applied NEFTune noise embeddings to fine-tuning. This has been proven to improve model performances for instrcution fine-tuning. Additionally, it underwent fine-tuning on a combination of publicly available and synthetic datasets through the use of Direct Preference Optimization (DPO). Interestingly, we observed an uplift in performance on the MT Bench when the intrinsic alignment of these datasets was eliminated, resulting in a more effective assistant model.

Developed by : Wonchul Kim (Lotte Data Communication AI Technical Team)

Hardware and Software

• Hardware: We utilized an A100x8 * 1 for training our model
• Training Factors: We fine-tuned this model using a combination of the DeepSpeed library and the HuggingFace TRL Trainer / HuggingFace Accelerate

Base Model : beomi/llama-2-koen-13b

Training Data

The LDCC-Instruct-Llama-2-ko-13B model was trained with publicly accessible Korean/English data sources. For its fine-tuning, we utilized other public data and underwent some processing and refinement.

We did not incorporate any client data owned by Lotte Data Communication.

Prompt Template

### Prompt:
{instruction}

### Answer:
{output}

License

LICENSE.txt