OwenArli/ArliAI-Llama-3-8B-Dolfin-v1.0

Based on Meta-Llama-3-8b-Instruct, and is governed by Meta Llama 3 License agreement: https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct

Base model: https://huggingface.co/failspy/Meta-Llama-3-8B-Instruct-abliterated-v3

SFT fine tune of Meta Llama 3 8B Instruct Abliterated v3 by Failspy using an improved Dolphin and WizardLM dataset intended to remove GPT-isms and make the model follow instructions more exactly while paying attention to details better.

Since it is based on the Abliterated version of Llama 3 8B Instruct it should naturally not refuse to answer in the first place and this fine tuning should make it comply even better.

Best practices:

• Be precise and explain what you want the model to do. It has less base "personality" than the OG model but it will act however you tell it to.
• This model works best with system prompts that tells it that it is the character, instead of telling it to act as a character.

Training:

• Full 8192 sequence length
• Training duration is around 2.5 days on an RTX 4090
• 1 epoch training with a massive dataset for minimized repetition sickness.
• Using 4-bit loading and Qlora 64-rank 64-alpha resulting in ~2% trainable weights.

Llama 3 Instruct format:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{{ system_prompt }}<|eot_id|><|start_header_id|>user<|end_header_id|>

{{ user_message_1 }}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

{{ model_answer_1 }}<|eot_id|><|start_header_id|>user<|end_header_id|>

{{ user_message_2 }}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Quants:

FP16: https://huggingface.co/OwenArli/ArliAI-Llama-3-8B-Dolfin-v1.0

GGUF: https://huggingface.co/OwenArli/ArliAI-Llama-3-8B-Dolfin-v1.0-GGUF