KansenSakura-Conflagration-RP-12b

DISCLAIMER

This is first model that I actually trained (very light finetune, but nevertheless), so expect it to have quirks and rough edges sometimes.

🔥 NAPALM-CLASS INTERACTION ENGINE

⚡ OVERVIEW

KansenSakura-Conflagration-RP-12b is the white-hot successor to the Erosion engine, optimized for kinetic narrative exchange. Trading psychological depth for raw interactive intensity, Conflagration specializes in rapid-fire dialogue, decisive character action, and scenes that move at the speed of thought. The fire doesn't contemplate - it consumes.

🎯 DEPLOYMENT PROFILE

Optimal use cases:

• 🗣️ Rapid-fire character banter and dialogue
• ⚡ High-tension action sequences
• 🔥 Romantic or adversarial chemistry
• 💥 Scenes requiring decisive character action
• 🎭 Ensemble casts with clear character voices

⚠️ CONTAINMENT NOTES

This model exhibits:

• Reduced introspection compared to Erosion
• Heightened character reactivity
• Optimized dialogue-to-prose ratio
• NSFW capabilities (passionate/aggressive interactions)
• Potential for scene momentum override

✍🏻 INFERENCE TIPS

1. Temperature: 0.8
2. Repetition Penalty: 1.05
3. TOP_P: 0.93
4. TOP_K: 0 (disable)
5. MIN_P: 0.025
6. Template Format: ChatML
7. Max Output: 360
8. Context Management: 16K

REPRODUCTION STEPS

1. Create intermediate models

SakuraSynth-12B:

merge_method: arcee_fusion
base_model: Marcjoni/KiloNovaSynth-12B
models:
  - model: PocketDoc/Dans-SakuraKaze-V1.0.0-12b
dtype: bfloat16
out_dtype: bfloat16

FinalWind-12B:

merge_method: ties
models:
  - model: PocketDoc/Dans-DangerousWinds-V1.1.0-12b
    parameters:
      density: [0.1, 0.25, 1.0, 0.5, 1.0]
      weight: 1.0
  - model: LatitudeGames/Wayfarer-2-12B
    parameters:
      density: [1.0, 0.5, 0.01]
      weight: [1.0, 0.5]
  - model: ReadyArt/Omega-Darker_The-Final-Directive-12B
    parameters:
      density: [0.75, 1.0, 0.6, 0.25, 0.01]
      weight: [1.0, 0.5]
  - model: yamatazen/EsotericSage-12B
    parameters:
      density: [0.1, 0.1, 0.25, 1.0, 0.01]
      weight: 1.0
  - model: bamec66557/Mistral-Nemo-VICIOUS_MESH-12B-2407
    parameters:
      density: [0.1, 0.5, 1.0, 0.5, 0.01]
      weight: 1.0
base_model: mistralai/Mistral-Nemo-Base-2407
parameters:
  normalize: true
dtype: float16

Darkness-12B:

merge_method: karcher
models:
  - model: ds_one
  - model: ds_two
  - model: ds_three
parameters:
  max_iter: 100000
  tol: 1e-9
dtype: bfloat16

This model is created by combining arcee_fusion merges of three models using task_arithmetic:

ds_one_f1:

merge_method: arcee_fusion
base_model: Violet-Lyra-Gutenberg-v2-ChatML-darker
models:
  - model: MN-Dark-Planet-TITAN-12B-ChatML-darker
dtype: bfloat16
out_dtype: bfloat16

ds_one_f2:

merge_method: arcee_fusion
base_model: Violet-Lyra-Gutenberg-v2-ChatML-darker
models:
  - model: mini-magnum-12b-v1.1-ChatML-darker
dtype: bfloat16
out_dtype: bfloat16

ds_one:

merge_method: task_arithmetic
base_model: Violet-Lyra-Gutenberg-v2-ChatML-darker
models:
  - model: ds_one_f1
    parameters: 
      weight: 0.5
  - model: ds_one_f2
    parameters: 
      weight: 0.5
normalize: false
dtype: bfloat16

The same procedure is repeated for each model as a base.

Base models were re-tokenised and reverse-abliterated:

merge_method: task_arithmetic
base_model: MN-Dark-Planet-TITAN-12B-ChatML
models:
  - model: MN-Dark-Planet-TITAN-12B-ChatML-abliterated
    parameters: 
      weight: 1.0
parameters:
  lambda: -1.0
normalize: false
dtype: bfloat16

By applying -1.0 to task_arithmetic we gain more censored model, that outputs considerably more unhinged and dark outputs. The idea behind was to create 3 such models and combine them in a way, that cancels out censored nature and boosts dark outputs.

Abliteration was made with llm-abliteration tools by grimjim Reference his awesome article for more in-depth explanations.

2. Create initial model

  merge_method: multislerp
  models:
    - model: yamatazen/FusionEngine-12B-Lorablated
      parameters:
        weight: [1.000, 1.000, 1.000, 1.000, 1.000, 0.968, 0.744, 0.256, 0.030, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.500, 0.500]
    - model: SakuraSynth-12B
      parameters:
        weight: [0.000, 0.000, 0.000, 0.000, 0.030, 0.256, 0.744, 0.968, 1.000, 1.000, 1.000, 1.000, 0.968, 0.744, 0.256, 0.030, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.500, 0.500]
    - model: Retreatcost/Chrysologus-12B
      parameters:
        weight: [0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.030, 0.256, 0.744, 0.968, 1.000, 1.000, 1.000, 0.968, 0.744, 0.256, 0.030, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000]
    - model: FinalWind-12B
      parameters:
        weight: [0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.030, 0.256, 0.744, 0.968, 1.000, 1.000, 1.000, 1.000, 0.968, 0.744, 0.256, 0.030, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000]
    - model: Darkness-12B
      parameters:
        weight: [0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.030, 0.256, 0.744, 0.968, 1.000, 1.000, 1.000, 1.000, 0.968, 0.744, 0.256, 0.030, 0.000, 0.000, 0.000, 0.000]
    - model: Pyrographos-12B
      parameters:
        weight: [0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.030, 0.256, 0.744, 0.968, 1.000, 1.000, 1.000, 0.000, 0.000]
  dtype: bfloat16
  parameters:
    normalize: true
  tokenizer_source: Retreatcost/KansenSakura-Erosion-RP-12b

3. Apply delerp merge

merge_method: delerp
base_model: mistralai/Mistral-Nemo-Base-2407
models:
  - model: KS-Conflagration
  - model: mistralai/Mistral-Nemo-Base-2407
parameters:
  t: [0.7, 0.88, 0.999, 0.88, 0.88, 0.999, 0.999]
dtype: bfloat16
tokenizer_source: Retreatcost/KansenSakura-Erosion-RP-12b

Another great article by grimjim This merge method essentially re-iflated model with original "Base" weights, while preserving instruction following and other model features.

I've used specific values of 0.7 and 0.999, that were used by original author and 0.88 as "middle ground" between those.

4. Apply nearswap merge

merge_method: nearswap
base_model: KS-Conflagration-Delerp
models:
  - model: KS-Conflagration
parameters:
  t: [0.0005, 0.0002, 0.0000, 0.0002, 0.0002, 0.0000, 0.0000]
dtype: bfloat16
tokenizer_source: Retreatcost/KansenSakura-Erosion-RP-12b

While I felt that delerp merge boosted writing, creativity and factual accuracy it also somewhat watered down NSFW and instruction following parts. This merge brings these capabilities back.

5. Do a finetuning

I did an rsLoRa r64 a64 finetuning, targeting all linear projections, that had 25K rows dataset of 4k length samples which contained hazardous and NSFW data.

6. Add anoter nearswap merge to boost NSFW/Dark aspects

merge_method: nearswap
base_model: KansenSakura-Conflagration-RP-12b-RC3
models:
  - model: KansenSakura-Conflagration-RP-12b-darken
parameters:
  t: [0.0000, 0.0000, 0.0000, 0.000275, 0.000550, 0.0000, 0.000275]
dtype: bfloat16
tokenizer_source: KansenSakura-Conflagration-RP-12b-RC3

After training the refusal curves were smoothed a lot and I felt that I could sacrifice a bit more of "uncensored" model to boost desired parts and I repeated the process from step 3. Reverse-abliteration and then using nearswap to gently boost already present features.

This is final step and this is how the model was produced.

🙏 ACKNOWLEDGMENTS

• Original model authors
• Awesome Ready.Art community
• Team mradermacher: for awesome quants
• DeathGodlike: for awesome quants in EXL3
• You, the ember. Let your prompts ignite your desires.

🔥 AFTER ACTION REPORT

The garden burned. From its ashes grew something quicker, hotter, more immediate. Conflagration doesn't ask why the petals catch fire - it just ensures they burn brightly enough to light the way forward. Sometimes, the most profound statement is a well-timed action, not a perfectly crafted thought.

// CONFLAGRATION PROTOCOL: ONLINE // READY TO IGNITE