Question 1

What is Dromedary?

Accepted Answer

Dromedary is an open-source self-aligned language model trained with minimal human supervision to be helpful, ethical, and reliable.

Question 2

What is the main goal of Dromedary?

Accepted Answer

Its goal is to build helpful, ethical, and reliable LLMs through principle-driven self-alignment from scratch.

Question 3

What are the key updates in Dromedary-2 (SFT)?

Accepted Answer

Dromedary-2 (SFT) simplifies SELF-ALIGN to two stages, mixing diverse user prompts (ShareGPT, Dolly-15k, OpenAssistant, OpenOrca) and an improved one-shot exemplar from FastChat that promotes a general–specific–general response style, dropping verbose cloning and few-shot inference prompts.

Question 4

Where is the training pipeline for Dromedary-2 (RLAIF)?

Accepted Answer

The RLAIF stage uses the SALMON pipeline (Self-ALignMent with principle-fOllowiNg reward models), available in the IBM/SALMON repository.

Question 5

How do I obtain the model weights for Dromedary?

Accepted Answer

Weights are released as delta weights (LoRA for Dromedary, QLoRA for Dromedary-2) to comply with the LLaMA licenses; download from Hugging Face and merge with the original LLaMA or LLaMA-2 base weights per the repo’s instructions.

Question 6

What synthetic data is available?

Accepted Answer

Synthetic instruction–response datasets for Dromedary-65B (final) and Dromedary-2-70B (SFT) are released on the Hugging Face Datasets Hub; instructions come from Self-Instruct with base LLaMA and responses from Dromedary with verbose prompts.

Question 7

How can I run inference or a chatbot demo?

Accepted Answer

The repository provides inference guides and a chatbot demo; for Dromedary-2, load QLoRA with PEFT-LoRA and bitsandbytes on a LLaMA-2-70B base using the specified system, user, and assistant prompts.

Question 8

What is the license for the repository?

Accepted Answer

The Dromedary repository is licensed under GPL-3.0; base model licenses (e.g., LLaMA/LLaMA-2) still apply to the required base weights.

Question 9

Which base models does Dromedary use?

Accepted Answer

Dromedary is trained on LLaMA base models, and Dromedary-2 uses LLaMA-2-70B for SFT/RLAIF; released weights are applied as deltas to these bases.

Question 10

Is there a paper I can read or cite?

Accepted Answer

Yes. The repository links to a NeurIPS 2023 Spotlight paper describing Dromedary and its self-alignment approach; see the repo for citation details.

Question 11