Stable Beluga

Tülu 3 is an advanced instruction-following language model developed by the Allen Institute for AI (Ai2), designed to enhance capabilities in areas such as knowledge, reasoning, mathematics, coding, and safety. Built upon the Llama 3 Base, Tülu 3 employs a comprehensive four-stage post-training process: meticulous prompt curation and synthesis, supervised fine-tuning on a diverse set of prompts and completions, preference tuning using both off- and on-policy data, and a novel reinforcement learning approach to bolster specific skills with verifiable rewards. This open-source model distinguishes itself by providing full transparency, including access to training data, code, and evaluation tools, thereby closing the performance gap between open and proprietary fine-tuning methods. Evaluations indicate that Tülu 3 outperforms other open-weight models of similar size, such as Llama 3.1-Instruct and Qwen2.5-Instruct, across various benchmarks.

StableVicuna is the first large-scale open source chatbot trained via reinforced learning from human feedback (RHLF). StableVicuna is a further instruction fine tuned and RLHF trained version of Vicuna v0 13b, which is an instruction fine tuned LLaMA 13b model. In order to achieve StableVicuna’s strong performance, we utilize Vicuna as the base model and follow the typical three-stage RLHF pipeline outlined by Steinnon et al. and Ouyang et al. Concretely, we further train the base Vicuna model with supervised finetuning (SFT) using a mixture of three datasets: OpenAssistant Conversations Dataset (OASST1), a human-generated, human-annotated assistant-style conversation corpus comprising 161,443 messages distributed across 66,497 conversation trees, in 35 different languages; GPT4All Prompt Generations, a dataset of 437,605 prompts and responses generated by GPT-3.5 Turbo; And Alpaca, a dataset of 52,000 instructions and demonstrations generated by OpenAI's text-davinci-003.