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
Mostly an experiment trying to completely uncensor the model, doesn't seem to be near as good as the original in reasoning and knowledge. It is however pretty good for RP.
Training:
- 4096 sequence length, while the base model is 8192 sequence length. From testing it still performs the same 8192 context just fine.
- Training duration is around 3 days on an RTX 4090, using 4-bit loading and Qlora 64-rank 128-alpha resulting in ~2% trainable weights.
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:
GGUF: https://huggingface.co/mradermacher/Llama-3-8B-Cumulus-v0.1-GGUF
