Aethora-7B-V1
Creator: SteelSkull
About Aethora: Trained on 2 Full Epochs of Aethora-7b-V1 using Aether-V1.9 Dataset, Aethora is a model trained specifically for general use with a focus in RP/Story based on the 2.5mil row (around 1 billion tokens) Aether dataset.
Model Quants: Quants provided by: [N/A] .
Model Sources:
- Developed & Funded by: Steelskull
- Finetuned from model: Mistral-7B-Instruct-v0.2
- Finetuning Repository: Aether Dataset
- Model type: BF16
- License: A2
Finetune Information:
- Hardware Type: H100 x1
- Hours Used: 60-Hrs
- Cloud Provider: Runpod.io
- Compute Region: US-IL
Dataset Information:
- Version v1.9: Fixed an error where 'system' and 'tools' records were not being carried over to the final dataframe. Added an 'origins' record for dataset sources.
- Version 1.8.5: Removed missing conversations or starting messages that are empty, and selectively omitted certain phrases for coherence and relevance.
Datasets Used:
- grimulkan/bluemoon_Karen_cleaned
- Doctor-Shotgun/no-robots-sharegpt
- Locutusque/Hercules-v3.0
- jondurbin/airoboros-3.2
- openerotica/freedom-rp
- teknium/OpenHermes-2.5
- Doctor-Shotgun/capybara-sharegpt
- KaraKaraWitch/PIPPA-ShareGPT-formatted
- Locutusque/bagel-clean-v0.3-shuffled
- Locutusque/hyperion-v3.0
Dataset Summary (Processed / Removed):
- Total Objects Removed: 209074
- Deduplication Stats: Starting row count: 4738917, Final row count: 2673175, Rows removed: 2065742