v000000/L3.1-Storniitova-8B

Llama-3.1-Storniitova-8B

Storniitova-8B is a RP/Instruct model built on the foundation of Llama-3.1-SuperNova-Lite, which is distilled from the 405B parameter variant of Llama-3.1

By only changing the vector tasks, I attempt to retain the full 405B distillation while learning roleplaying capabilties.

(GGUF) mradermacher quants:

• GGUFs
• GGUFs imatrix

(GGUF) QuantFactory quants:

• GGUFs

---

merge

This is a merge of pre-trained language models created using mergekit and other proprietary tools.

Merge Details

Merge Method

This model was merged using the SLERP, Task_Arithmetic and NEARSWAP merge method.

Models Merged

The following models were included in the merge:

• v000000/L3.1-Niitorm-8B-t0.0001
• akjindal53244/Llama-3.1-Storm-8B
• arcee-ai/Llama-Spark
• arcee-ai/Llama-3.1-SuperNova-Lite
• v000000/L3.1-8B-RP-Test-003-Task_Arithmetic
• Sao10K/L3.1-8B-Niitama-v1.1 + grimjim/Llama-3-Instruct-abliteration-LoRA-8B
• v000000/L3.1-8B-RP-Test-002-Task_Arithmetic + grimjim/Llama-3-Instruct-abliteration-LoRA-8B

Recipe

The following YAML configuration was used to produce this model:

#Step1 - Add smarts to Niitama with alchemonaut's algorithm.

slices:
  - sources:
      - model: Sao10K/L3.1-8B-Niitama-v1.1+grimjim/Llama-3-Instruct-abliteration-LoRA-8B
        layer_range: [0, 32]
      - model: akjindal53244/Llama-3.1-Storm-8B
        layer_range: [0, 32]
merge_method: nearswap
base_model: Sao10K/L3.1-8B-Niitama-v1.1+grimjim/Llama-3-Instruct-abliteration-LoRA-8B
parameters:
  t:
    - value: 0.0001
dtype: bfloat16
out_type: float16

#Step 2 - Learn vectors onto Supernova 0.4(Niitorm)

models:
  - model: arcee-ai/Llama-3.1-SuperNova-Lite
    parameters:
      weight: 1.0
  - model: v000000/L3.1-Niitorm-8B-t0.0001
    parameters:
      weight: 0.4
merge_method: task_arithmetic
base_model: arcee-ai/Llama-3.1-SuperNova-Lite
parameters:
    normalize: false
dtype: float16

#Step 3 - Fully learn vectors onto Supernova 1.25(Niitorm)

models:
  - model: arcee-ai/Llama-3.1-SuperNova-Lite
    parameters:
      weight: 0.0
  - model: v000000/L3.1-Niitorm-8B-t0.0001
    parameters:
      weight: 1.25
merge_method: task_arithmetic
base_model: arcee-ai/Llama-3.1-SuperNova-Lite
parameters:
    normalize: false
dtype: float16

#Step 4 - Merge checkpoints and keep output/input Supernova heavy
#Merge with a triangular slerp from sophosympatheia.

models:
  - model: v000000/L3.1-8B-RP-Test-003-Task_Arithmetic
merge_method: slerp
base_model: v000000/L3.1-8B-RP-Test-002-Task_Arithmetic+grimjim/Llama-3-Instruct-abliteration-LoRA-8B
# This model needed some abliteration^
parameters:
  t:
    - value: [0, 0, 0.3, 0.4, 0.5, 0.6, 0.5, 0.4, 0.3, 0, 0]
dtype: float16

SLERP distribution used to smoothly blend the mostly Supernova base with the roleplay vectors:

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	28.06
IFEval (0-Shot)	78.17
BBH (3-Shot)	30.81
MATH Lvl 5 (4-Shot)	13.29
GPQA (0-shot)	5.26
MuSR (0-shot)	9.96
MMLU-PRO (5-shot)	30.84