👻 Geodesic Phantom 12B

This is a merge of pre-trained language models created using mergekit.

This was merged in 7 hours on a runpod A40 using an adaptive VRAM chunking script (based on measure.py by GrimJim)

WARNING:mergekit.graph:OOM at chunk 65536, reducing to 32768 (attempt 1, progress: 0/131075)
WARNING:mergekit.graph:OOM at chunk 32768, reducing to 16384 (attempt 2, progress: 0/131075)

[Karcher_Stock Audit] Layer: lm_head.weight
Stats: Cos(θ): 0.564 | t-factor: 0.8843 | Karcher Iters: 2960
  (Base)  mistralai--Mistral-Nemo-Instruct-2407               : █████                                              ( 11.57%)
  (Donor) Vortex5--Prototype-X-12b                            : ███████                                            ( 14.74%)
  (Donor) Vortex5--Stellar-Witch-12B                          : ███████                                            ( 14.74%)
  (Donor) Vortex5--Celestial-Queen-12B                        : ███████                                            ( 14.74%)
  (Donor) Vortex5--Moonlit-Mirage-12B                         : ███████                                            ( 14.74%)
  (Donor) Vortex5--Crimson-Constellation-12B                  : ███████                                            ( 14.74%)
  (Donor) Vortex5--Wicked-Nebula-12B                          : ███████                                            ( 14.74%)

The following patch was also required for this merge

karcher_stock Adaptive Tanh Soft-Clamp v11

# ── 11. Model Stock t factor with Adaptive Soft-Clamp ─────────────
        N = len(ws_2d)
        ct = cos_theta.unsqueeze(-1) if cos_theta.dim() > 0 else cos_theta
        
        # Raw Model Stock formula
        denom = 1.0 + (N - 1) * ct
        # Add a tiny epsilon to prevent literal division by zero
        t_raw = (N * ct) / denom.clamp(min=1e-6) 

        # --- BULLETPROOF TANH CLAMP ---
        # 1. Prevent negative infinity spikes (fallback to base model)
        t_clamped_bottom = torch.clamp(t_raw, min=0.0)
        
        # 2. Smoothly asymptote positive spikes to L (Maximum allowed t-factor)
        L = 1.5 
        excess = torch.clamp(t_clamped_bottom - 1.0, min=0.0)
        t_soft_top = 1.0 + (L - 1.0) * torch.tanh(excess / (L - 1.0))
        
        # 3. Apply: If t <= 1.0, use exact math. If t > 1.0, use soft curve.
        t = torch.where(t_clamped_bottom <= 1.0, t_clamped_bottom, t_soft_top)
        # ------------------------------

Example of the clamp preventing merge corruption

Merge Details

Merge Method

This model was merged using the karcher_stock merge method using /workspace/models/mistralai--Mistral-Nemo-Instruct-2407 as a base.

Models Merged

The following models were included in the merge:

• /workspace/models/Vortex5--Wicked-Nebula-12B
• /workspace/models/Vortex5--Celestial-Queen-12B
• /workspace/models/Vortex5--Moonlit-Mirage-12B
• /workspace/models/Vortex5--Stellar-Witch-12B
• /workspace/models/Vortex5--Prototype-X-12b
• /workspace/models/Vortex5--Crimson-Constellation-12B

Configuration

The following YAML configuration was used to produce this model:

architecture: MistralForCausalLM
base_model: /workspace/models/mistralai--Mistral-Nemo-Instruct-2407
models:
  - model: /workspace/models/Vortex5--Prototype-X-12b
  - model: /workspace/models/Vortex5--Celestial-Queen-12B
  - model: /workspace/models/Vortex5--Wicked-Nebula-12B
  - model: /workspace/models/Vortex5--Stellar-Witch-12B
  - model: /workspace/models/Vortex5--Moonlit-Mirage-12B
  - model: /workspace/models/Vortex5--Crimson-Constellation-12B
merge_method: karcher_stock # v8
parameters:  
  filter_wise: true
  max_iter: 10000
  min_iter: 1000
  tol: 1.0e-11
dtype: float32
out_dtype: bfloat16
tokenizer:
  source: union
chat_template: auto
name: 👻 Geodesic Phantom 12B