README.md

metadata

base_model:
  - NousResearch/Meta-Llama-3-8B-Instruct
library_name: transformers
tags:
  - mergekit
  - merge
  - llama
  - conversational
license: llama3

L3-Hecate-8B-v1.2

About:

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

Recommended Samplers:

Temperature - 1.0
TFS - 0.7
Smoothing Factor - 0.3
Smoothing Curve - 1.1
Repetition Penalty - 1.08

Merge Method

This model was merged a series of model stock, followed by ExPO. It uses a mix of roleplay models to improve performance.

Configuration

The following YAML configuration was used to produce this model:

---
# Concise-Mopey
models:
  - model: Salesforce/LLaMA-3-8B-SFR-Iterative-DPO-Concise-R
    parameters:
      weight: 1.0
  - model: failspy/Llama-3-8B-Instruct-MopeyMule
    parameters:
      weight: 1.0
merge_method: task_arithmetic
base_model: NousResearch/Meta-Llama-3-8B-Instruct
parameters:
  normalize: false
dtype: float32
vocab_type: bpe
name: Concise-Mopey

---
# Mopey RP Mix
models:
  - model: Concise-Mopey+Azazelle/Llama-3-Sunfall-8b-lora
  - model: Concise-Mopey+Azazelle/Llama-3-8B-Abomination-LORA
  - model: Concise-Mopey+Azazelle/llama3-8b-hikikomori-v0.4
  - model: Concise-Mopey+Azazelle/Llama-3-Instruct-LiPPA-LoRA-8B
  - model: Concise-Mopey+Azazelle/BlueMoon_Llama3
  - model: Concise-Mopey+Azazelle/Llama3_RP_ORPO_LoRA
  - model: Concise-Mopey+mpasila/Llama-3-LimaRP-Instruct-LoRA-8B
  - model: Concise-Mopey+Azazelle/Llama-3-LongStory-LORA
merge_method: model_stock
base_model: failspy/Llama-3-8B-Instruct-MopeyMule
dtype: float32
vocab_type: bpe
name: mopey_rp

---
models:
  - model: Nitral-AI/Hathor_Tahsin-L3-8B-v0.85
  - model: Sao10K/L3-8B-Tamamo-v1
  - model: Sao10K/L3-8B-Niitama-v1
  - model: Hastagaras/Jamet-8B-L3-MK.V-Blackroot
  - model: nothingiisreal/L3-8B-Celeste-v1
  - model: Jellywibble/lora_120k_pref_data_ep2
  - model: Nitral-AI/Hathor_Stable-v0.2-L3-8B
  - model: mopey_rp
merge_method: model_stock
base_model: NousResearch/Meta-Llama-3-8B-Instruct
dtype: float32
vocab_type: bpe
name: hq_rp

---
# ExPO
models:
  - model: hq_rp
    parameters:
      weight:
        - filter: mlp
          value: 1.15
        - filter: self_attn
          value: 1.025
        - value: 1.0
merge_method: task_arithmetic
base_model: NousResearch/Meta-Llama-3-8B-Instruct
parameters:
  normalize: false
dtype: float32
vocab_type: bpe