Llama3 Typhoon v1.5 (scb10x) LLM

Llama3 Typhoon v1.5 (scb10x) LLM 

Typhoon-1.5 models come in 8B and 72B sizes

These models are built on the 8B Llama3 and 72B Qwen base models, 8B weights released under the Meta Llama 3 Community License; 72B weights released under the Tongyi Qianwen License

Performance

To gain insight into Typhoon’s performance, we evaluated it using multiple-choice exam:

Language & Knowledge Capabilities: We assessed Typhoon on multiple-choice question answering datasets, including ThaiExam, M3Exam, and MMLU. The ThaiExam dataset was sourced from standard examinations in Thailand, including ONET, TGAT, TPAT, and A-Level. M3Exam is a benchmark for Southeast Asian countries, including Thailand. MMLU is a standard benchmark for language models in English.



Typhoon-1.5X is an eXperimental model designed for application use cases, featuring improved capabilities in Retrieval-Augmented Generation (RAG), constrained generation, and reasoning in order to achieve competitive performance in instruction following and better human alignment in Thai.


Llama3-Typhoon v1.5x 70B 4bit GGUF Demo on Colab



https://huggingface.co/Adun
Download for Hugging Face

https://huggingface.co/Adun/llama-3-typhoon-v1.5x-70b-instruct-gguf



Use NVIDIA A100 40GB GPU on Colab

This 70B model has 81 layers and size 42.5GB. But NVIDIA A100 GPU has only 40GB RAM.
So, We cannot load all 81 layers. Then We set model_kwargs={"n_gpu_layers": 70}.









Inference Model





Install llama-cpp-python




https://opentyphoon.ai/

Typhoon API

Access the Typhoon Instruction Tuned model through our new API service.
Available for free during our open beta.
https://huggingface.co/Adun

Adun Nantakaew อดุลย์ นันทะแก้ว

LINE : adunnan


ความคิดเห็น

โพสต์ยอดนิยมจากบล็อกนี้

Eval Llama v3.1 8B ,70B Model with Thai Language

LLAMA-CPP-PYTHON on RTX4060 GPU