Graig Alpha: Artificial Grain Intelligence

electron271/graig-alpha is a 4 billion parameter language model developed by electron271, building upon the "tuxsentience" series by @GrainWare. This model is uniquely focused on "artificial grain intelligence," aiming to provide insights and capabilities within the agricultural domain.

Key Features & Advancements

• Specialized Domain: Unlike general-purpose LLMs, Graig Alpha is fine-tuned for agricultural applications, leveraging expertise from grain farmers.
• Local Fine-tuning: The model was fine-tuned entirely locally using Unsloth on an AMD RX 9070 XT, benefiting from recent advancements in AMD support within Unsloth.
• Extended Context Window: Graig Alpha supports a native context length of up to 262,144 tokens, though 32,768 tokens are recommended for reduced RAM usage, enabling the processing of extensive agricultural data.

Recommended Usage Settings

To optimize performance and mitigate potential issues, the following settings are recommended:

• temperature = 0.6
• top_p = 0.95
• top_k = 20
• min_p = 0.00
• presence_penalty = 0.0 to 2.0 (e.g., 1.0 to reduce repetitions)

Considerations

While newer Graig models are less prone to offensive statements, users should be aware that such occurrences can still happen due to the inherent unpredictability of LLMs, especially if settings are not configured correctly or prompts are not appropriate. For public deployments, implementing content filters like llmcordplus is advised.