allenai/tulu-v1-llama2-70b

TEXT GENERATIONConcurrency Cost:4Model Size:69BQuant:FP8Ctx Length:32kPublished:Nov 12, 2023Architecture:Transformer0.0K Cold

allenai/tulu-v1-llama2-70b is a 69 billion parameter instruction-tuned language model developed by AllenAI, fine-tuned from Meta's Llama 2. This model is designed to act as a helpful assistant, trained on a diverse mix of publicly available, synthetic, and human-created datasets. It specializes in generating helpful responses based on its v1 Tulu data mixture, making it suitable for assistant-style conversational AI applications.

Loading preview...

Tulu 1 Llama 2 70B: An Instruction-Tuned Assistant Model

Tulu 1 Llama 2 70B is a 69 billion parameter language model developed by AllenAI, fine-tuned from the meta-llama/Llama-2-70b-hf base model. It is part of the Tulu series, which focuses on creating helpful assistant models through instruction and RLHF tuning.

Key Capabilities

  • Helpful Assistant: Trained to act as a conversational assistant, providing informative and relevant responses.
  • Diverse Training Data: Fine-tuned on the Tulu v1 data mixture, which includes a blend of publicly available, synthetic, and human-generated datasets.
  • Instruction Following: Designed to follow specific instructions, making it suitable for various task-oriented applications.
  • Primarily English: Optimized for English language processing.

Intended Uses & Limitations

This model is best suited for applications requiring a helpful, instruction-following chatbot. Users should be aware that, unlike some other models, Tulu models have not undergone extensive alignment for safe completions within an RLHF phase or deployed with in-the-loop filtering. Consequently, it may produce problematic outputs if specifically prompted to do so. The model expects inputs formatted with <|user|> and <|assistant|> tags, with a newline after <|assistant|> for optimal generation quality.

For more details, refer to the associated research paper: Camels in a Changing Climate: Enhancing LM Adaptation with Tulu 2.