krevas/LDCC-Instruct-Llama-2-ko-13B-v4.1.12

TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:4kLicense:cc-by-nc-4.0Architecture:Transformer Open Weights Cold

LDCC-Instruct-Llama-2-ko-13B-v4.1.12 is a 13 billion parameter instruction-tuned causal language model developed by Lotte Data Communication, based on Meta's Llama 2 architecture. This model is fine-tuned for dialogue use cases, leveraging supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align with human preferences for helpfulness and safety. It is optimized for generating text responses in an assistant-like chat format, building upon the Llama 2's 4k context length and 2.0T pretraining tokens.

Loading preview...

Model Overview

LDCC-Instruct-Llama-2-ko-13B-v4.1.12 is an instruction-tuned large language model developed by Lotte Data Communication, built upon Meta's Llama 2 13B architecture. This model has been fine-tuned using a combination of the DeepSpeed library and HuggingFace Trainer/Accelerate on A100x8 hardware.

Key Characteristics

  • Base Model: Utilizes the Llama 2 13B pretrained model, which features an optimized transformer architecture and a 4k token context length.
  • Fine-tuning: Optimized for dialogue use cases through supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF), aiming for improved helpfulness and safety.
  • Prompt Template: Designed to work with a specific prompt format: ### Prompt:\n{instruction}\n\n### Answer:\n{output}.

Intended Use Cases

This model is primarily intended for assistant-like chat applications. While the base Llama 2 models are generally for commercial and research use in English, this specific instruction-tuned version by Lotte Data Communication suggests an adaptation for specific dialogue tasks, potentially with a focus on Korean language capabilities given the model name's 'ko' suffix, though the README primarily details the English Llama 2 base.

Limitations

As with all LLMs, the model may produce inaccurate, biased, or objectionable responses. Developers are advised to perform safety testing tailored to their specific applications. The base Llama 2 model's intended use is primarily English, and its performance in other languages is not guaranteed without further specific fine-tuning.

Finetunes
2 models