Daemontatox/Llama-3.3-8B-Instruct
Daemontatox/Llama-3.3-8B-Instruct is an 8 billion parameter instruction-tuned causal language model, derived from an unreleased Llama 3.3 8B version from Meta's Llama API. This model offers improved performance over Llama 3.1 8B Instruct, particularly in reasoning tasks as indicated by IFEval and GPQA benchmarks. It is notable for its unique origin, being extracted and redistributed from Meta's API, and is suitable for general instruction-following applications.
Loading preview...
Llama 3.3 8B Instruct Overview
Daemontatox/Llama-3.3-8B-Instruct is an 8 billion parameter instruction-tuned model, representing a previously unreleased Llama 3.3 8B version from Meta's Llama API. This model was uniquely extracted by finetuning through the API and then subtracting the adapter to recover the base model. It offers an opportunity to access a version of Llama 3.3 8B that was not publicly available.
Key Capabilities & Performance
- Improved Reasoning: Benchmarks show notable improvements over Llama 3.1 8B Instruct.
- IFEval: Achieves 81.95 (compared to 78.2 for Llama 3.1 8B Instruct).
- GPQA Diamond: Scores 37.0 (compared to 29.3 for Llama 3.1 8B Instruct).
- Context Length: The original downloaded model has an 8k context length, though a community-modified version with a 128k context length (using Llama 3.3 70B RoPE config) shows slightly better benchmark results.
- Unique Origin: This model's existence and redistribution are based on a novel method of extraction from Meta's finetuning API, making it distinct from other Llama releases.
Should I use this for my use case?
- Early Access to Llama 3.3: If you are interested in exploring an early, unreleased version of Llama 3.3 8B, this model provides that opportunity.
- General Instruction Following: Its instruction-tuned nature makes it suitable for a wide range of general-purpose conversational and instruction-based AI tasks.
- Benchmarking & Research: Researchers and developers can use this model to compare performance against other Llama 3.x variants, especially given its unique origin and benchmark improvements.