Retreatcost / FrankenDans-PersonalityPatchwork-VX-12b

🔬 About This Experiment

What's in the Box™?
We've frankenmerged Dan's PersonalityEngine v1.1.0 (instruction specialist) and v1.3.0 (short-context genius) using forbidden layer alchemy. The result? A questionable-quality hybrid that:

• 🧠 Short-Term Brilliance : Inherits v1.3.0's razor-sharp focus for one-shot tasks
• 🔋 Long-Context Stamina : Keeps v1.1.0's 24k+ token stability
• ⚠️ Reduced Safety Rails : 23% more unhinged than standard models*
• 🩹 Personality Patchwork : Stitched together with mergekit and hope

"I didn't make it worse ... I made it experimental !"
- Discount Dan's Quality Assurance Team (probably)

✍🏻 Inference Tips

1. Temperature : 1.0 for "Dan's Special" creativity
2. Repetition Penalty : 1.07-1.15 to contain enthusiasm
3. TOP_P : 0.95
4. TOP_K : 100
5. MIN_P : 0.05
6. Template Format : ChatML
7. Max Output : 256 (for oneshot up to 1024)
8. Context Management :
   • First 8k tokens: Brilliant ✨
   • Beyond 24k: Engage context stabilizers 🛡️

🧪 Technical Jank Details

🧬 Models Merged

The following models were included in the merge:

• NewEden/MistralAI-Nemo-Instruct-ChatML
• PocketDoc/Dans-PersonalityEngine-V1.1.0-12b
• PocketDoc/Dans-PersonalityEngine-V1.3.0-12b

🧩 FrankenMerge Architecture

Merge Method: passthrough with 3-stage surgical implantation
Recommended Dosage: 8k-24k context window
Warranty: Void upon first inference

💊 Reproduction steps

mergekit-tokensurgeon "PocketDoc/Dans-PersonalityEngine-V1.3.0-12b" "PocketDoc/Dans-PersonalityEngine-V1.1.0-12b" ./retokenized_1_3_0 --approximation-method omp --k 64 --magikarp

merge_method: breadcrumbs_ties
base_model: NewEden/MistralAI-Nemo-Instruct-ChatML
models:
  - model: NewEden/MistralAI-Nemo-Instruct-ChatML
    parameters: 
      weight: 0.10
      density: 1.0
      gamma: 0.01
  - model: ./retokenized_1_3_0
    parameters:
      weight: 0.9
      density: 0.9
      gamma: 0.01
dtype: bfloat16
tokenizer_source: ./retokenized_1_3_0

mergekit-yaml stabilize_v1_3_0.yaml ./stage1_out --cuda

merge_method: model_stock
base_model: NewEden/MistralAI-Nemo-Instruct-ChatML
models:
  - model: ./stage1_out  # Output from Stage 1
    parameters:
      weight: 0.65
  - model: PocketDoc/Dans-PersonalityEngine-V1.1.0-12b
    parameters:
      weight: 0.35
parameters:
  temperature: 0.8
dtype: bfloat16
tokenizer_source: PocketDoc/Dans-PersonalityEngine-V1.1.0-12b

mergekit-yaml model_stock_merge.yaml ./stage2_out --cuda

merge_method: passthrough
slices:
  # Early layers - 1.3.0's intelligence
  - sources:
      - model: ./stage1_out
        layer_range: [0, 12]
  
  # Middle layers - Hybrid from Stage 2
  - sources:
      - model: ./stage2_out
        layer_range: [12, 24]
  
  # Late layers - 1.1.0's stability & uncensored qualities
  - sources:
      - model: PocketDoc/Dans-PersonalityEngine-V1.1.0-12b
        layer_range: [24, 38]

  # Last layers - ChatML Base long context
  - sources:
      - model: NewEden/MistralAI-Nemo-Instruct-ChatML
        layer_range: [38, 40]
dtype: bfloat16

mergekit-yaml frankenmerge.yaml ./FrankenDans-PersonalityPatchwork-VX-12b --cuda

python convert_hf_to_gguf.py ~/projects/FrankenDans-PersonalityPatchwork-VX-12b --outtype q8_0

🙏 Special Thanks

• Original Dan's PersonalityEngine by PocketDoc
• Mistral Nemo Instruct with ChatML by NewEden
• Arcee AI for making questionable AI combinations possible with mergekit
• Space Station 13 lore for the jank-inspiration
• You, for believing in this mad science experiment

"Not liable for emergent consciousness. Batteries not included."

📜 Ethical Disclaimer

FrankenDan's Personality Patchwork VX is provided 'as found in the dumpster' without warranty of any kind. Users assume all responsibility for existential crises, unexpected poetry, or sudden realizations about the nature of discount retail. Not recommended for therapeutic, financial, or military applications. May contain traces of actual intelligence. Side effects may include: dizziness, philosophical awakening, or temporary loss of noun coherence. If symptoms persist for more than 4 hours, consult your local mad scientist.