![Lyra](https://huggingface.co/Sao10K/MN-12B-Lyra-v4/resolve/main/lyra.png) Mistral-NeMo-12B-Lyra-v4, a variation of [Lyra-v4a1](https://huggingface.co/Sao10K/MN-12B-Lyra-v4a1), layered over [Lyra-v3](https://huggingface.co/Sao10K/MN-12B-Lyra-v3), which was built on top of [Lyra-v2a2](https://huggingface.co/Sao10K/MN-12B-Lyra-v2a2), which itself was built upon [Lyra-v2a1](https://huggingface.co/Sao10K/MN-12B-Lyra-v2a1). # Automatic Approval If you agree, please place future merges / derivatives under cc-by-nc-4.0 license. ty --- # Model Versioning ``` [See Previous Models] | Lyra-v4a1 | ------------> Lyra-v4 [Seperate RL Step targeting Instruct and Coherency over Base Nemo instead of SFT First, Result is Merged with Lyra-v4a1, fixes most quant-based issues. Somehow.] ``` # This uses ChatML, or any of its variants which were included in previous versions. ``` <|im_start|>system This is the system prompt.<|im_end|> <|im_start|>user Instructions placed here.<|im_end|> <|im_start|>assistant The model's response will be here.<|im_end|> -------------------------------------------------- [INST]system This is another system prompt.[/INST] [INST]user Your instructions placed here.[/INST] [INST]assistant The model's response will be here.[/INST] ``` # Recommended Samplers: ``` Temperature: 0.6 - 1 # Make sure min_p is set before Temperature in Sampler Orders min_p: 0.1 - 0.2 # Crucial for NeMo ``` # Recommended Stopping Strings: ``` <|im_end|> [/INST] ``` # Notes \- I think I fixed the extra token stuff some users seem to be facing, while retaining everything else? It's some error alright.
\- If you're using XML tags, you may see weird malformed stopping strings. Just add them to your current list. and move on.
\- Its pretty nice, imo. I've been messing around with it a lot.
\- Make sure the ChatML template is correct, I think there's some issues with the one used in SillyTavern which might cause improper replies? --- Reup due to issues I faced on my page, crashing every time i go in here. works fine now? so yay? ``` Changes: Weights are the same, literally. go and check the checksum. Credit to ArliAI for the model tokenizer configs and json I yoinked, to fix the tokenizer mess (in v3) I made, hence why config.json had their model name. Same stuff applies ```