#llama3 #silytavern #multimodal #llama3

[!TIP] I apologize for disrupting your experience.
My upload speeds have been cooked and unstable lately.
If you want and you are able to...
You can support my various endeavors by clicking here (Ko-fi).

[!IMPORTANT] Relevant:
These quants have been done after the fixes from llama.cpp/pull/6920 have been merged.
Use KoboldCpp version 1.64 or higher, make sure you're up-to-date.

[!WARNING] Compatible SillyTavern presets recommended by the authors are hosted in the repository files or here (Virt's Roleplay Presets - my personal recommendation).
Use the latest version of KoboldCpp. Use the provided presets for testing.
Feedback and support for the Authors is always welcome.
If there are any issues or questions let me know.

[!NOTE] For 8GB VRAM GPUs, I recommend the Q4_K_M-imat (4.89 BPW) quant for up to 12288 context sizes.

Cat-llama3-instruct

Presented by:

• Dataset Builder: Dr. Kal'tsit (Kat)
• Trainer/Funding: SteelSkull
• Facilitator: Potatooff

70b variant of the model (Trained by Dr. Kal'tsit, Posted by Turboderp), Please check it out!

About:

Cat-llama3-instruct is a llama 3 8b finetuned model focusing on system prompt fidelity, helpfulness and character engagement. The model aims to respect system prompt to an extreme degree, provide helpful information regardless of situations, and offer maximum character immersion (Role Play) in given scenes.

Specific Aims:

• System Instruction fidelity
• Chain of Thought (COT)
• Character immersion
• Helpfulness for biosciences and general science

Methods

Dataset Preparation

Huggingface dataset containing instruction-response pairs was systematically pulled. A GPT model was trained on high quality and helpful responses exclusively to serve as a standard model.

(Fig1. Huggingface dataset population distribution and filtration for each component)

The dataset was further filtered for length and COT responses:

(Fig 2. COT responses are all >50 tokens single turn)

Health-related data was also pulled from Chat Doctor, favoring detailed and step-by-step diagnosis:

(Fig 3. Health related tasks are all >100 tokens, peaking at 450 tokens single turn)

Model Training

1*a100 for 6 days, 4 epochs

Usage:

Prompt format: llama 3

• Settings/Context/Instruct (Provided by @Anthonyg5005)
• Cat 8B Gen Settings
• Cat 8b Context
• Cat 8b Instruct

Quants:

• [bartowski/llama-3-cat-8b-instruct-v1-GGUF]

 <|begin_of_text|>Below is a conversation between an user and an AI assistant. The ai assistant gives response regarding user's questions. USER: user: Given the task definition and input, reply with output. In this task, you are given a question and an answer, you would be asked to create the sentence based on the Question-Answer provided. It should be contained within the Question-Answer provided.
[object Object]

[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]
[object Object]

1. Double-check the sentence for proper grammar and punctuation, and make necessary adjustments if needed.
   Final Output:
   In humans, protist diseases are caused by protozoa.< / s ><|end_of_text|>

Showcases:

The model does COT in grey, calculated response in black; note that this behaviour is done through system card instruction to showcase the systemcard fidelity, it is not fine tuned into the model.

(Fig 4. Showcasing model doing COT to solve difficult tasks, extend and enrich its own answers)