Lambent/Arsenic-Shahrazad-12B-v4

Experimenting with trying to provide Arsenic with a lightweight update to her knowledge base via QLoRA, to give her some idea what's happened since 2024. She absorbed some of it, not other parts.

Also shifted the chat template to somewhat more resemble Devstral's, to start training with system prompts, and cold-started her on a few hermes-agent traces with the reasoning stripped.

Her style and self-perception shifted a lot more modern overall, understandably, so likely to revisit with original and on-distribution DPO for a style update soon.

arsenic-ku-v3-merged

This is a merge of pre-trained language models created using mergekit.

Merge Details

Merge Method

This model was merged using the Karcher Mean merge method.

Models Merged

The following models were included in the merge:

• —../trainingmodels/outputs/bakedv3_seed3
• —../trainingmodels/outputs/bakedv3_seed2
• —../trainingmodels/outputs/bakedv3_seed1

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: ../training_models/outputs/baked_v3_seed1
  - model: ../Projects/training_models/outputs/baked_v3_seed2
  - model: ../training_models/outputs/baked_v3_seed3
merge_method: karcher
parameters:
  normalize: true
  int8_mask: true
tokenizer_source: ../training_models/outputs/baked_v3_seed1
dtype: bfloat16