Merge Method

SLERP (Spherical Linear Interpolation) via mergekit.

Tankie's influence is deliberately contained to the middle layers (12–28) at a weight of 0.20. Early layers (0–12) and late layers (28–40) remain pure Rocinante-X — preserving embedding coherence and output quality while letting a hint of Tankie's character seep into the model's deeper reasoning.

yaml
Copy code
slices:
  - sources:
      - model: TheDrummer/Rocinante-X-12B-v1
        layer_range: [0, 12]

  - sources:
      - model: TheDrummer/Rocinante-X-12B-v1
        layer_range: [12, 28]
        parameters:
          weight: 0.80
      - model: WokeAI/Tankie-DPE-12B-SFT-v2
        layer_range: [12, 28]
        parameters:
          weight: 0.20

  - sources:
      - model: TheDrummer/Rocinante-X-12B-v1
        layer_range: [28, 40]

merge_method: slerp
base_model: TheDrummer/Rocinante-X-12B-v1

parameters:
  t:
    - value: 0.20

dtype: bfloat16

---

⚠️ Warnings

• Experimental. Behavior is untested and may be unpredictable.
• Ideological bias. Tankie-DPE is explicitly trained toward Marxist-Leninist positions. These views may surface in unexpected contexts — even with a neutral system prompt.
• Not for production. Personal research only.

---

Credits

Full credit and respect to the creators of the source models:

• TheDrummer for Rocinante-X-12B-v1 — genuinely one of the best RP models in its class.
• WokeAI for Tankie-DPE-12B-SFT-v2 — a fascinating and unusual training experiment, whatever one thinks of its politics.
• And once again, a thousand thanks to team mradermacher for providing the excellent quants.

---

License

Inherits from base models. Please refer to the respective model cards for full license details.
Shared for personal research use only.