Llama 3 8B Instruct

Llama 3 8B Instruct is the 8-billion-parameter, instruction-tuned model from Meta's first Llama 3 release, dated April 18, 2024 on its Hugging Face model card. It is an autoregressive transformer with Grouped-Query Attention, tuned with supervised fine-tuning and reinforcement learning from human feedback for assistant-style chat. Meta's card notes that a newer version, Llama 3.1 8B Instruct, is available.

Key features

• Training scale: pretrained on more than 15 trillion tokens of publicly available data, with a knowledge cutoff of March 2023 for the 8B model.
• Instruction tuning: fine-tuned on publicly available instruction datasets plus over 10 million human-annotated examples. Meta says neither pretraining nor fine-tuning data includes Meta user data.
• Context: an 8K-token context window.
• Safety work: Meta reports red-teaming and tuning to reduce unnecessary refusals, and recommends Llama Guard and other safety tooling in deployments.
• Meta-reported results: MMLU 68.4 (5-shot), GPQA 34.2, HumanEval 62.2, GSM-8K 79.6 and MATH 30.0 for the instruction-tuned 8B model, according to the card. These are Meta's own numbers.

Model specs

Use cases

• Local and self-hosted assistant chat in English, the card's primary language. Developers may fine-tune for other languages if they follow the license and Acceptable Use Policy.
• Fine-tuning experiments and research baselines at the 8B scale.
• Legacy deployments that were built on Llama 3 and need a stable reference.

Getting started

1. Accept the Meta Llama 3 Community License on the model card.
2. Load the model with Hugging Face Transformers, use the original llama3 codebase, or use a llama.cpp-compatible quantization.
3. Add Llama Guard or similar safety tooling before production use.
4. For new work, evaluate Llama 3.1 8B Instruct first, since the card itself points to it as the newer version.

FAQ

What is Llama 3 8B Instruct's context length?

8K tokens, according to Meta's model card.

When was Llama 3 8B Instruct released?

April 18, 2024.

What license does it use?

The custom Meta Llama 3 Community License, which requires agreeing to share your Hugging Face information before access.

What is the newer version?

Meta's card points to Llama 3.1 8B Instruct, which has its own entry at Llama 3.1 8B Instruct.

Alternatives

• Llama 3.1 8B Instruct: the newer version of this model.
• Llama 3 70B Instruct: the larger Llama 3 model, with a December 2023 knowledge cutoff.
• Llama Guard 3: Meta's safety classifier for Llama deployments.

Conclusion

Llama 3 8B Instruct is a compact, widely deployed open-weight chat model, and a handy baseline. Read the official model card for license terms and usage, and browse the models category for newer options.