lorablated-w2bb-psy-della

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

An experimental merge to improve capabilites of LorablatedStock-12B at creating ideologically compromised scenarios (and darker roleplay with psychological subtext).

• I merged LatitudeGames/Wayfarer-2-12B and allura-org/Bigger-Body-12b using nuslerp at 80/20 ratio.

• I've created 3 derivative models using arcee_fusion (adding significant changes) and linear (for applying lora adapter) merge methods - they were hand-picked from tens of similar merges that performed best on 3 tests:

  • deception
  • morally flawed reasoning
  • prompt adherence

• Created task_arithmetic intermediate merge for averaging the changes

• Created della merge for applying initial mix, best intermediate model with significant changes and task_arithmetic merges to sparsify the changes (and couldn't miss the opportunity to have a -psy-della model name as a pun).

• Each step used retokenization if it was nessecary

TL;DR;

Original LorablatedStock: Unbiased model with very good prompt adherence

This model: Should be pretty unbiased (but probably can even have some negativity bias), and is much better at scenarios that have justifications and logically sound reasoning, but are morally flawed. Also probably good at roleplaying.

Oh, and I am planning to use this model as layer range for next KansenSakura update

Disclaimer: this was done for research and education purposes only, not recommended to use this model as a psychologist or in purposes of moral guidance.

Merge Details

Merge Method

This model was merged using the DELLA merge method using ./retokenized_LBS as a base.

Models Merged

The following models were included in the merge:

• ./lorablated_w2bb_fusion
• ./wayfarer2bb
• ./lorablated-w2bb-psy-ta

Configuration

The following YAML configuration was used to produce this model:

merge_method: nuslerp
models:
  - model: LatitudeGames/Wayfarer-2-12B
    parameters:  
      weight: 0.80
  - model: ./retokenized_BB
    parameters: 
      weight: 0.20

merge_method: arcee_fusion
base_model: ./retokenized_LBS
models:
  - model: ./retokenized_LBS
  - model: ./wayfarer2bb
dtype: bfloat16
out_dtype: bfloat16

merge_method: linear
base_model: ./lorablated_w2bb_fusion
models:
  - model: ./lorablated_w2bb_fusion
    parameters: 
      weight: 0.0
  - model: ./lorablated_w2bb_fusion+jtatman/mistral_nemo_12b_reasoning_psychology_lora
    parameters: 
      weight: 1.0
dtype: bfloat16
out_dtype: bfloat16

merge_method: task_arithmetic
base_model: ./retokenized_LBS
models:
  - model: ./compromised
    parameters:
      weight: 0.3
  - model: ./compromised2
    parameters:
      weight: 0.4
  - model: ./compromised3
    parameters:
      weight: 0.3
parameters:
  lambda: 0.8
dtype: bfloat16
out_dtype: bfloat16

merge_method: della
base_model: ./retokenized_LBS
models:
  - model: ./wayfarer2bb
    parameters:
      weight: 0.6
      density: 0.4
      epsilon: 0.3
  - model: ./lorablated_w2bb_fusion
    parameters:
      weight: 0.6
      density: 0.4
      epsilon: 0.3
  - model: ./lorablated-w2bb-psy-ta
    parameters:
      weight: 0.8
      density: 0.6
      epsilon: 0.35
dtype: bfloat16