license: llama3
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
This is by far the most completely uncensored Llama 3 8b instruct model. It will literally never refuse anything. So as a reminder, with great power comes great responsibility.
In terms of reasoning and intelligence, this model is probably worse than the OG model because of the decensoring. However, if you have issues with refusals then this will be superior just because it will not refuse.
Will soon have quants uploaded here on HF and have it up on our site https://awanllm.com for anyone to try.
OpenLLM Benchmark:
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.
- Added DPO fine tuning aside from a more curated dataset for this v0.2 model.
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/AwanLLM/Awanllm-Llama-3-8B-Cumulus-v0.2
GGUF: https://huggingface.co/AwanLLM/Awanllm-Llama-3-8B-Cumulus-v0.2-GGUF