Hastagaras/Halu-8B-Llama3-Blackroot

Hastagaras/Halu-8B-Llama3-Blackroot: An Experimental Merged Model

Hastagaras/Halu-8B-Llama3-Blackroot is an experimental 8 billion parameter language model with an 8192-token context length, created by Hastagaras. It was developed using the Model Stock merge method, combining Hastagaras/Halu-8B-Llama3-v0.3 with two LoRAs from Blackroot: Llama-3-LongStory-LORA and Llama-3-8B-Abomination-LORA. The developer notes that the model exhibits "very strange" and "human-like" characteristics, possibly due to the human-sourced data used in Blackroot's LoRAs.

Key Characteristics & Performance

• Architecture: Based on the Llama 3 family, merged from Hastagaras/Halu-8B-Llama3-v0.3 and Blackroot's LoRAs.
• Context Length: Supports an 8192-token context window.
• Performance: Achieves an average score of 69.78 on the Open LLM Leaderboard. Notable scores include:
  • HellaSwag (10-Shot): 84.55
  • Winogrande (5-Shot): 79.48
  • GSM8k (5-Shot): 70.51
• Unique Output: Described as generating "very human-like" actions and narrations.

Important Considerations

• Experimental Status: This model is explicitly labeled as experimental and has not been extensively tested or evaluated. Its performance characteristics are largely unknown.
• Safety Warning: The model is noted to be "very unsafe in certain parts," particularly in roleplay (RP) scenarios, and may generate harmful, biased, or inappropriate content. Users are advised to exercise extreme caution and use it at their own risk.

GGUF Availability

Quantized GGUF versions are available, including an IMATRIX GGUF by Lewdiculous and a STATIC GGUF by mradermacher.