VerwaltungsAnthologie_clear_simbad_7B

VerwaltungsAnthologie_clear_simbad_7B

This model is used as an intermediate model for future merges. This is a merge of 4 pre-trained language models based upon Mistral-7B-v0.1 created using mergekit.

In combination with DiscoLM_German_7b_v1 this 'clear'-model is the 'base' model to build the successor of my first 'VA_Disco_7B': VerwaltungsAnthologie_Disco_7B

Merge Details

Merge Method

This model was merged using the DARE TIES merge method using mistralai/Mistral-7B-v0.1 as a base.

Models Merged

The following models were included in the merge:

• hiig-piai/simba-01d-ftb
• DRXD1000/Phoenix
• OpenPipe/mistral-ft-optimized-1227
• VAGOsolutions/SauerkrautLM-7b-LaserChat

Explanations of used models

• Main focus of this "clear" model is the hiig-piai/simba-01d (strangley without a 'v') by the Humboldt Institute for Internet and Society has built this model for "simplified language" (Leichte Sprache in german).
• The DRXD1000/Phoenix got finetuned with many german texts of law – it can even "hallucinate" almost perfect URL of the official archive of german laws: [Gesetze im Internet]](https://www.gesetze-im-internet.de/)
• OpenPipe/mistral-ft-optimized-1227 performed best using mixed languages in combination with mistralai/Mistral-7B-v0.1 as base model.
• VAGOsolutions/SauerkrautLM-7b-LaserChat has a wider range of colloquial german language.
• mistralai/Mistral-7B-v0.1 is the base model – funny but true – only using OpenPipe/mistral-ft-optimized-1227 as base model is not as good as combining both.

Configuration

The following YAML configuration was used to produce this model:

# works but never stops
models:
  - model: mistralai/Mistral-7B-v0.1
    # No parameters necessary for base model
  - model: VAGOsolutions/SauerkrautLM-7b-LaserChat
    parameters:
      density: 0.53
      weight: 0.15
  - model: hiig-piai/simba-01d-ftb
    parameters:
      density: 0.53
      weight: 0.55
  - model: DRXD1000/Phoenix
    parameters:
      density: 0.53
      weight: 0.15
  - model: OpenPipe/mistral-ft-optimized-1227
    parameters:
      density: 0.53
      weight: 0.15
merge_method: dare_ties
base_model: mistralai/Mistral-7B-v0.1
parameters:
  int8_mask: true
dtype: bfloat16
name: VerwaltungsAnthologie_clear_simbad_7B