---
base_model: []
library_name: transformers
tags:
- mergekit
- merge

---
> [!CAUTION]
> <span style="color:red; font-weight:bold">⚠️ Warning:</span> This merge produces BROKEN output and is not recommended to download. The tensorguard method needs revision.
>

# 💂 TensorGuard-Prototype-24B-v1

This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit).

## Merge Details
### Merge Method

This model was merged using the [TensorGuard](https://arxiv.org/abs/2506.01631v2) merge method.

### Models Merged

The following models were included in the merge:
* /workspace/Naphula--BeaverAI_Fallen-Mistral-Small-3.1-24B-v1e_textonly
* /workspace/TheDrummer--Magidonia-24B-v4.3
* /workspace/TheDrummer--Precog-24B-v1
* /workspace/TheDrummer--Cydonia-24B-v4.3

### Configuration

The following YAML configuration was used to produce this model:

```yaml
architecture: MistralForCausalLM
models:
  - model: /workspace/Naphula--BeaverAI_Fallen-Mistral-Small-3.1-24B-v1e_textonly
  ## 2506 ##
  - model: /workspace/TheDrummer--Cydonia-24B-v4.3
  ## 2509 ##
  - model: /workspace/TheDrummer--Precog-24B-v1
  - model: /workspace/TheDrummer--Magidonia-24B-v4.3
merge_method: tensorguard # https://arxiv.org/abs/2506.01631v2
parameters:
  noise_epsilon: 0.01                    # Noise magnitude for perturbations
  num_perturbations: 30                  # Number of perturbation iterations (paper default)
  noise_strategies: "adversarial,structural,low_freq,high_freq,gaussian" # All noise strategies from paper
  similarity_metric: "frobenius"         # Distance metric: frobenius, spectral, euclidean, cosine
  normalize_weights: true                # Normalize weights to sum to 1
  random_seed: 420                       # Seed for reproducible results
  pca_components: 8                      # PCA components for dimensionality reduction
  use_higher_order_stats: true           # Compute skewness and kurtosis (expensive)
  use_spectral_features: true            # Compute spectral norm features (very expensive)
tokenizer:
  source: union
chat_template: auto
dtype: float32
out_dtype: bfloat16
name: 💂 Tensorguard-24B-v1
```