metadata
license: llama3
tags:
- moe
language:
- en
8bpw/h8 exl2 quantization of xxx777xxxASD/L3-ChaoticSoliloquy-v2-4x8B-test using default exllamav2 calibration dataset.
ORIGINAL CARD:
(Maybe i'll change the waifu picture later.)
TRY 1.5 or 1.0 INSTEAD AND CHECK IF THEY WORK BETTER, I DIDN'T TEST THIS VERSION BEFORE PUBLISHING
Experimental RP-oriented MoE, the idea was to get a model that would be equal to or better than the Mixtral 8x7B and it's finetunes in RP/ERP tasks.
The model has totally 25B parameters, of which ~13B are active.
Please feedback me if it's more stable than the previous version
Llama 3 ChaoticSoliloquy-v2-4x8B test
base_model: L3_ChaosMaid_8B
gate_mode: random
dtype: bfloat16
experts_per_token: 2
experts:
- source_model: ChaoticNeutrals_Poppy_Porpoise-0.72-L3-8B
- source_model: L3_ChaosMaid_8B
- source_model: openlynn_Llama-3-Soliloquy-8B-v2
- source_model: Sao10K_L3-Solana-8B-v1
Models used
- ChaoticNeutrals/Poppy_Porpoise-0.72-L3-8B
- jeiku/Chaos_RP_l3_8B
- NeverSleep/Llama-3-Lumimaid-8B-v0.1
- openlynn/Llama-3-Soliloquy-8B-v2
- Sao10K/L3-Solana-8B-v1
Difference
- Update from ChaoticNeutrals/Poppy_Porpoise-v0.7-L3-8B to ChaoticNeutrals/Poppy_Porpoise-0.72-L3-8B
- Update from openlynn/Llama-3-Soliloquy-8B to openlynn/Llama-3-Soliloquy-8B-v2
- Change - NeverSleep/Llama-3-Lumimaid-8B-v0.1 to L3-ChaosMaid-8B
L3 ChaosMaid-8B
models:
- model: jeiku_Chaos_RP_l3_8B
# No parameters necessary for base model
- model: NeverSleep_Llama-3-Lumimaid-8B-v0.1
parameters:
density: 0.5
weight: 0.5
merge_method: dare_ties
base_model: jeiku_Chaos_RP_l3_8B
parameters:
int8_mask: true
dtype: bfloat16
Vision
