Hugging Face Overview
Overview
BAAI/Infinity-Instruct-7M-Gen-Llama3_1-8B is an 8 billion parameter instruction-tuned model developed by the Beijing Academy of Artificial Intelligence (BAAI). It is built upon the Llama 3.1 architecture and is notable for being fine-tuned exclusively through supervised instruction tuning, without the use of reinforcement learning from human feedback (RLHF).
Key Capabilities & Training
- Instruction Following: The model is fine-tuned on the extensive Infinity-Instruct-7M and Infinity-Instruct-Gen datasets, which are million-level instruction datasets.
- Foundational Ability Enhancement: An initial phase involved tuning on Infinity-Instruct-7M to improve foundational abilities, including math and code, before further fine-tuning for chat.
- Performance: Benchmarks indicate strong performance, with the model achieving 33.9 on AlpacaEval 2.0, which is competitive with or superior to Llama-3-8B-Instruct and Llama-3.1-8B-Instruct, and showing favorable results against GPT-4 in some metrics.
- Training Efficiency: The training process leveraged techniques from FlagScale to concatenate multiple training samples and apply acceleration, reducing training costs.
Use Cases
This model is suitable for general instruction-following tasks and conversational AI applications. Its competitive performance on benchmarks like AlpacaEval 2.0 suggests its utility in scenarios requiring robust response generation without the overhead of RLHF.