Hugging Face's logo

Triangle104
/
L3.1-Aglow-Vulca-v0.1-8B-Q6_K-GGUF

Transformers
GGUF
mergekit
Merge
roleplay
RP
storytelling
llama-cpp
gguf-my-repo
Inference Endpoints
conversational
Model card Files Community
L3.1-Aglow-Vulca-v0.1-8B-Q6_K-GGUF / README.md
Triangle104's picture
Triangle104
Update README.md
34998d6 verified
preview code
|
raw
history blame
6.04 kB
metadata
base_model: kromeurus/L3.1-Aglow-Vulca-v0.1-8B
library_name: transformers
license: cc-by-nc-4.0
tags:
  - mergekit
  - merge
  - roleplay
  - RP
  - storytelling
  - llama-cpp
  - gguf-my-repo

Triangle104/L3.1-Aglow-Vulca-v0.1-8B-Q6_K-GGUF

This model was converted to GGUF format from kromeurus/L3.1-Aglow-Vulca-v0.1-8B using llama.cpp via the ggml.ai's GGUF-my-repo space. Refer to the original model card for more details on the model.

Model Details & Recommended Settings

This is a story telling first model that is proficient in narrative driven RP. Does best with straightforward instructions. Any wishy-washy language will confuse it. As per usual with any of my models with Formax in it, it's pretty sensitive to instructs so choose your words wisely.

Once going though, it's able to generate detailed and human-ish outputs with lots of personality depending on the information given. Has a habit of matching the style and format of the input; style, spacing, grammar, etc. Can interweave details from the character persona, chat history, and user persona (if there is one) to create unique interactions and plot points. Leans more or less positive naturally but can be flipped if prompted correctly.

Being a Llama 3.1 model, it's still subject to the normal pros and cons of L3/L3.1 but I'd like to think I tamed some of it. Keep the temp on the lower end since there is a low chance it might freak out. If it does, swipe/regen the chat or delete the afflicted output and try again.

Rec. Settings:

Template: Llama 3 Token Count: 128k Max Temperature: 1.2 Min P: 0.1 Repeat Penalty: 1.05 Repeat Penalty Tokens: 256

Merge Theory

Where to begin. The general though process was still the roughly same as the Ablaze, making one very smart model and another more creative focused model. This time, I merged Formax and RPmax in separately instead of doing one merge since they have different focuses.

'Apollobulk' is the smarts, having the storytelling capabilities from badger writer, instruct following from Formax (duh) and the smarts of Super Nova. Apollo 0.4 was use as an RP temper to keep the overall model aligned with RP. Apollo 2.0 wasn't used as it skewed the merge too far towards inconsistent narratives.

'Reshape' is the creative end, taking some inspo from the Ablaze's creative center. First created 'Darkened' as the main influence over the final writing style of Aglow. Poppy Moonfall C had the personality I was looking for but the smarts (though not important was still necessary) so the other three were added to round out it's overall capabilities while being very creative. Plopping that atop RPmax (For excellent unique RP interactions), BRAG (serious recall), and Natsumura (a great Storytelling/RP base) and model stock it, you get a really solid model on its own.

Slap the two components together in a simple gradient dare_linear merge and boom; this unit of an 8B model. As of writing and releasing this model, mergekit is fucked for me (one of its dependencies has broken L3 merging) so I can't test any other methods atm. If there is a better final merge method, I'll be uploading a v0.2 once the bug is fixed.

This time around, everything was done with DavidAU's High Quality method, merging with float32 at all steps. Made a significant difference in nuanced understanding of text. Config

models: - model: Locutusque/Apollo-0.4-Llama-3.1-8B - model: maldv/badger-writer-llama-3-8b - model: ArliAI/Llama-3.1-8B-ArliAI-Formax-v1.0 base_model: arcee-ai/Llama-3.1-SuperNova-Lite parameters: int8_mask: true merge_method: model_stock dtype: float32 tokenizer_source: base name: apollobulk

models: - model: v000000/L3-8B-Poppy-Moonfall-C - model: Casual-Autopsy/Jamet-L3-Stheno-BlackOasis-8B - model: SicariusSicariiStuff/Dusk_Rainbow base_model: ResplendentAI/Rawr_Llama3_8B parameters: int8_mask: true merge_method: model_stock dtype: float32 tokenizer_source: base name: darkened

models: - model: darkened - model: ArliAI/Llama-3.1-8B-ArliAI-RPMax-v1.1 - model: maximalists/BRAG-Llama-3.1-8b-v0.1 base_model: tohur/natsumura-storytelling-rp-1.0-llama-3.1-8b parameters: int8_mask: true merge_method: model_stock dtype: float32 tokenizer_source: base name: reshape

models:

  • model: reshape parameters: weight: [0.1, 0.9]
  • model: apollobulk parameters: weight: [0.9, 0.1] base_model: reshape tokenizer_source: base parameters: normalize: false int8_mask: true merge_method: dare_linear dtype: float32 name: vulca

Use with llama.cpp

Install llama.cpp through brew (works on Mac and Linux)

brew install llama.cpp

Invoke the llama.cpp server or the CLI.

CLI: 

llama-cli --hf-repo Triangle104/L3.1-Aglow-Vulca-v0.1-8B-Q6_K-GGUF --hf-file l3.1-aglow-vulca-v0.1-8b-q6_k.gguf -p "The meaning to life and the universe is"

Server: 

llama-server --hf-repo Triangle104/L3.1-Aglow-Vulca-v0.1-8B-Q6_K-GGUF --hf-file l3.1-aglow-vulca-v0.1-8b-q6_k.gguf -c 2048

Note: You can also use this checkpoint directly through the usage steps listed in the Llama.cpp repo as well.

Step 1: Clone llama.cpp from GitHub.

git clone https://github.com/ggerganov/llama.cpp

Step 2: Move into the llama.cpp folder and build it with LLAMA_CURL=1 flag along with other hardware-specific flags (for ex: LLAMA_CUDA=1 for Nvidia GPUs on Linux).

cd llama.cpp && LLAMA_CURL=1 make

Step 3: Run inference through the main binary.

./llama-cli --hf-repo Triangle104/L3.1-Aglow-Vulca-v0.1-8B-Q6_K-GGUF --hf-file l3.1-aglow-vulca-v0.1-8b-q6_k.gguf -p "The meaning to life and the universe is"

or 

./llama-server --hf-repo Triangle104/L3.1-Aglow-Vulca-v0.1-8B-Q6_K-GGUF --hf-file l3.1-aglow-vulca-v0.1-8b-q6_k.gguf -c 2048