Sao blesses us with another Lyra MN finetune. V2 is my daily driver currently so I'm aching to get this.
This is the 6bpw EXL2 Quant of this model. You can find the original model here.
For the 8bpw version, go here
For the 4bpw version, go here

Ungated. Thanks for the patience!

Mistral-NeMo-12B-Lyra-v3, built on top of Lyra-v2a2, which itself was built upon Lyra-v2a1.

Model Versioning

Lyra-v1 [Merge of Custom Roleplay & Instruct Trains, on Different Formats]
  |
  | [Additional SFT on 10% of Previous Data, Mixed]
  v
Lyra-v2a1 
  |
  | [Low Rank SFT Step + Tokenizer Diddling]
  v
Lyra-v2a2
  |
  | [RL Step Performed on Multiturn Sets, Magpie-style Responses by Lyra-v2a2 for Rejected Data]
  v
Lyra-v3

This uses a custom ChatML-style prompting Format!

-> What can go wrong?

[INST]system
This is the system prompt.[/INST]
[INST]user
Instructions placed here.[/INST]
[INST]assistant
The model's response will be here.[/INST]

Why this? I had used the wrong configs by accident. The format was meant for an 8B pruned NeMo train, instead it went to this. Oops.

Recommended Samplers:

Temperature: 0.7 - 1.2
min_p: 0.1 - 0.2 # Crucial for NeMo

Recommended Stopping Strings:

<|im_end|>
</s>

Blame messed up Training Configs, oops?

Training Metrics:

- Trained on 4xH100 SXM for 6 Hours.
- Trained for 2 Epochs.
- Effective Global Batch Size: 128.
- Dataset Used: A custom, cleaned mix of Stheno-v3.4's Dataset, focused mainly on multiturn.

Extras

Image Source: AI-Generated with FLUX.1 Dev.

have a nice day.