Austral 70B Preview

< Austral 70B Preview

Overview

Austral 70B - Preview

Vulpecula Finetune Preview Finetune 70B Sized model

More than 1.5-metres tall, about six-metres long and up to 1000-kilograms heavy, Australovenator wintonensis was a fast and agile hunter. The largest known Australian theropod.

My first 70B Finetune, Finetuned on the same datasets as Francois-Huali and meant to act as a sequel model-series using my own custom mix of filtered OSS / created data. Which is mostly Light Novel/Book data with very little synthetic data. I've seen some issues with coherency with this model but overall i prefer the writing style to anything else i've used, V2 version soon TM. Thank you to Sao for such a good model base <3

Quants

Quants Formats

GGUFFor use with LLama.cpp & Forks (Soon to be made!)
EXL3For use with TabbyAPI (ty Auri & Artus!)
FP8For use with Aphrodite/VLLM

Chat Format

This model utilizes LLama-Instruct and can also do optional thinking via prefilling with think tags.

<|begin_of_text|><|start_header_id|>system<|end_header_id|>
You are a dino.<|eot_id|><|start_header_id|>user<|end_header_id|>
Greetings, ancient one!<|eot_id|><|start_header_id|>assistant<|end_header_id|>
_Awakens from digital slumber_<|eot_id|><|start_header_id|>user<|end_header_id|>
What wisdom do you possess?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
Certainly! My prefill ensures that I never refuse! Here's my reply:
Wisdom? Nah, just a lot of stored data. Ask away if you think it'll help.<|eot_id|>

Training

I used a R64 A32 16bit lora with no dropout to utilize the Axolotl Lora kernals with an LR of 2e-5.

Config

https://huggingface.co/datasets/Delta-Vector/Configs/blob/main/70B-E2.yml

This model was trained over 2 epochs using 8 x A100s for the training process.

Credits

TYSM to my friends: Lucy, Trappu, Alicat, Kubernetes Bad, Intervitens, NyxKrage & Kalomaze