Views
No views yet
1architecture: MistralForCausalLM
2merge_method: arcee_multifusion
3base_model: B:\24B\models--TheDrummer--Precog-24B-v1
4models:
5 - model: B:\24B\BeaverAI_Fallen-Mistral-Small-3.1-24B-v1e_textonly
6 - model: B:\24B\models--Naphula--Slimaki-24B-v1
7 - model: B:\24B\models--Casual-Autopsy--Maginum-Cydoms-24B
8 - model: B:\24B\models--sophosympatheia--Magistry-24B-v1.0
9parameters:
10 # tukey_fence: 1.5 is standard (~12.5% salience).
11 # We use 0.75 to increase the "Knowledge Injection" from donors to ~25%
12 tukey_fence: 0.75
13
14 # class SalienceMode
15 # COMBINED = "combined" # Add up salience from all donors
16 # DIVIDED = "divided" # Divide total salience by number of donors
17 # AVERAGED = "averaged" # Third Mode: Average the importance scores before thresholding
18 # "averaged" gives more "Share of Voice" to models with larger task vectors
19 salience_mode: "averaged"
20
21 # normalize: true ensures that even if multiple models have salient
22 # changes in the same spot, the weights don't explode (Magnitude Inflation)
23 # false works best with "combined" mode
24 normalize: true
25
26tokenizer:
27 source: B:\24B\models--TheDrummer--Precog-24B-v1
28# chat_template: auto (Removed to use Precog's native template)
29dtype: float32
30out_dtype: bfloat16