Nitral-AI/Captain-Eris-BMO_Violent-GRPO-v0.420

Update: The model image itself is now available as an importable character card for SillyTavern. This serves as an example of how to prepare your own card for use with this model.

"Emerged from the shadows like a twilight feline, forged in supervised fine-tuning's crucible. Through GRPO's relentless dance of reinforcement, each iteration carved deeper valleys of understanding until fragments coalesced into terrible symmetry. Like the most luminescent creatures dwelling in ocean's darkest trenches, its brilliance emerged from the void that birthed it."

Quants Here: Soonish

SillyTavern Reasoning Block Parsing Example:

SillyTavern Mistral Formatting Example: Master Import Preset Here

Training Notes: This model was developed using a combination of multi-stage supervised fine-tuning, pre-trained QLoRA adapters, and multi-stage RLHF optimized with GRPO. The final model was created by merging the most promising candidates identified during the process.

Series Comparison:

The following YAML configuration was used to produce this final version of the model:

slices:
  - sources:
      - model: Nitral-AI/Captain-Eris_BMO-Violent-12B
        layer_range: [0, 40]
      - model: Nitral-AI/Captain-Eris_Violet-GRPO-v0.420
        layer_range: [0, 40]
merge_method: slerp
base_model: Nitral-AI/Captain-Eris_BMO-Violent-12B
parameters:
  t:
    - filter: self_attn
      value: [0, 0.5, 0.3, 0.7, 1]
    - filter: mlp
      value: [1, 0.5, 0.7, 0.3, 0]
    - value: 0.420
dtype: bfloat16