Mergedonia-Suite-24B-v1

[!NOTE] <span style="color:red; font-weight:bold">⚠️ Note:</span> This project is slowly under construction. See this page for more info.

🚀 Mergedonia Suite 24B v1

This is test suite to see how various merge methods affects a model's ability at metrics and benchmarks.

Instead of gemma 2 finetunes, mistral 24B was used. The following models are featured:

• Cydonia v4.3
• Magidonia v4.3
• Precog v1
• Fallen Mistral v1e
• Rivermind

For methods that require a base_model, the most optimal is either anthracite-core/Mistral-Small-3.2-24B-Instruct-2506-Text-Only or Darkhn/Magistral-2509-24B-Text-Only, depending on what elements you want to emphasize.

Mergedonia is set to use 2509 as base_model due to it producing more consistent results overall than 2506 in my latest experiments.

The top 3 drummer finetunes plus the most uncensored (Fallen Mistral) as well as a random outlier (Rivermind) was included to see how each method reacts to:

• censorship
• intelligence
• style
• positivity
• context retention

So far there are 16 merges planned. These might compete in a bracket on challonge in the 16 > 8 > 4 > 2 > 1 format.

Update: More methods are planned to be included. The outline below is obsolete but is being updated soon.

Regular methods featured:

• arcee_fusion (2 stages)
• slerp (2 stages)
• multislerp
• model_stock
• karcher
• SCE
• della
• dare_ties
• ties

Custom:

• cvs
• brf
• magic
• chiral_qhe
• flux
• scream
• arcane

list randomizer recommended these for round 1:

• scream vs model_stock
• arcane vs arcee_fusion
• karcher vs ties
• slerp vs cvs
• magic vs della
• SCE vs multislerp
• brf vs chiral_qhe
• dare_ties vs flux

New methods are being added.