OpenAssistant/llama2-70b-oasst-sft-v10

TEXT GENERATIONConcurrency Cost:4Model Size:69BQuant:FP8Ctx Length:32kPublished:Aug 21, 2023License:llama2Architecture:Transformer0.1K Open Weights Cold

OpenAssistant/llama2-70b-oasst-sft-v10 is a 69 billion parameter causal decoder-only transformer language model, fine-tuned by Open-Assistant from Meta's Llama2 70B. It was trained in two stages, first on synthetic instructions and coding tasks, then polished with human demonstrations from open-assistant.io. This model is designed for instruction-following and general conversational AI, with enhanced compatibility for chat inference applications through its use of the OpenAI ChatML prompt format.

Loading preview...

Overview

OpenAssistant/llama2-70b-oasst-sft-v10 is a 69 billion parameter instruction-tuned language model, building upon Meta's Llama2 70B. Developed by Open-Assistant, this model underwent a two-stage fine-tuning process. The initial stage involved training on a diverse mix of synthetic instructions and coding tasks, while the second "polishing" stage utilized high-quality human demonstrations collected from open-assistant.io up to July 2023.

Key Features & Training

  • Base Model: Fine-tuned from meta-llama/Llama-2-70b.
  • Multilingual Support: Primarily English, with limited capabilities in German, Spanish, French, Italian, Portuguese, Polish, Dutch, Romanian, Czech, and Swedish.
  • Prompt Format: Adopts the OpenAI ChatML standard for improved compatibility with chat applications.
  • Training Data: Stage 1 included datasets like megacode2, orca-chat, dolly15k_multilingual, and oa_leet10k. Stage 2 focused on oasst_export (top-1 human demonstrations).
  • Context Length: Supports a sequence length of 4096 tokens during training.

Use Cases

This model is well-suited for general instruction-following tasks and conversational AI applications, particularly those benefiting from the ChatML prompt format. Developers should perform safety testing and tuning for specific deployments, as with all LLMs.