Llama 3.3 8B Instruct Overview

This model, allura-forge/Llama-3.3-8B-Instruct, is presented as an official 8 billion parameter instruction-tuned version of Meta's Llama 3.3. It was uniquely obtained by extracting the base model from Meta's Llama API's finetuning service, which initially offered a Llama 3.3 8B variant not publicly available elsewhere.

Key Capabilities & Performance

• Improved Instruction Following: Benchmarks indicate superior performance compared to Llama 3.1 8B Instruct.
  • IFEval (instruction following): 81.95 (vs 78.2 for Llama 3.1 8B Instruct)
  • GPQA Diamond: 37.0 (vs 29.3 for Llama 3.1 8B Instruct)
• Context Length: The original downloaded version has an 8k context length, but a variant configured for 128k context (using Llama 3.3 70B's RoPE config) shows slightly better benchmark results.
• Authenticity: Despite some internal metadata pointing to Llama 3, the model exhibits distinct stylistic and knowledge differences from Llama 3 and 3.1, aligning with the Llama API's Llama 3.3 8B offering.

Good For

• Developers seeking an instruction-tuned model with improved performance over Llama 3.1 8B Instruct.
• Experimentation with a potentially official, yet uniquely sourced, Llama 3.3 8B model.
• General instruction-following applications where the enhanced benchmark scores are beneficial.