Archangel SFT Llama 13B Overview

ContextualAI's archangel_sft_llama13b is a 13 billion parameter model built on the Llama architecture, specifically fine-tuned for instruction following. It leverages a Supervised Fine-Tuning (SFT) loss function and incorporates alignment data from SHP, Anthropic HH, and Open Assistant datasets to enhance its conversational capabilities.

Key Capabilities & Features

• Instruction Following: Optimized for responding to user prompts in a structured, assistant-like manner.
• TuluV2 Prompting Format: Requires a specific <|user|> and <|assistant|> turn-based format for optimal interaction, with the human speaking first.
• Context Length: Supports a context window of 4096 tokens.
• Conditional SFT: Models trained with conditional SFT include special <|good|> and <|bad|> tokens in their tokenizers, allowing for controlled generation by appending these to prompts.
• Automatic BOS Token: Automatically adds a beginning-of-sequence (BOS) token during tokenization, simplifying prompt preparation.

Alignment and Research

This model is part of Contextual AI's Human-Centered Loss Functions (HALOs) research, which focuses on improving LLM alignment. Further details on the training methodology and research can be found in their code repository and technical paper.