5 bits/bpw quantization of Venus-103b-v1.1 to be used on exllamav2.
Calibration dataset was a cleaned Pippa dataset (https://huggingface.co/datasets/royallab/PIPPA-cleaned), same as used as on the original model card.
You can use the measurement.json from there to do your own quant sizes
Original model card
Venus 103b - version 1.1
Model Details
- A result of interleaving layers of Sao10K/Euryale-1.3-L2-70B, migtissera/SynthIA-70B-v1.5, and Xwin-LM/Xwin-LM-70B-V0.1 using mergekit.
- The resulting model has 120 layers and 103 billion parameters.
- See mergekit-config.yml for details on the merge method used.
- See the
exl2-*branches for exllama2 quantizations. The 5.65 bpw quant should fit in 80GB VRAM, and the 3.35 bpw quant should fit in 48GB VRAM. - Inspired by Goliath-120b
Warning: This model will produce NSFW content!
Results
- Seems to be more "talkative" than Venus-103b-v1.0 (i.e characters speakmore often in roleplays)
- Sometimes struggles to pay attention to small details in the scenes
- Prose seems pretty creative and more logical than Venus-120b-v1.0
- Downloads last month
- 8
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.