L3-Aethora-15B
The Skullery Presents L3-Aethora-15B.
Creator: Steelskull
Dataset: Aether-Lite-V1.2
Trained: 4 x A100 for 15 hours Using RsLora and DORA
About L3-Aethora-15B:
L3 = Llama3 L3-Aethora-15B was crafted through using the abilteration method to adjust model responses. The model's refusal is inhibited, focusing on yielding more compliant and facilitative dialogue interactions. It then underwent a modified DUS (Depth Up Scale) merge (originally used by @Elinas) by using passthrough merge to create a 15b model, with specific adjustments (zeroing) to 'o_proj' and 'down_proj', enhancing its efficiency and reducing perplexity. This created AbL3In-15b.
AbL3In-15b was then trained for 4 epochs using Rslora & DORA training methods on the Aether-Lite-V1.2 dataset, containing ~82000 high quality samples, designed to strike a fine balance between creativity, slop, and intelligence at about a 60/40 split
This model is trained on the L3 prompt format.
Quants:
Dataset Summary: (Filtered)
Filtered Phrases: GPTslop, Claudism's
- mrfakename/Pure-Dove-ShareGPT: Processed 3707, Removed 150
- mrfakename/Capybara-ShareGPT: Processed 13412, Removed 2594
- jondurbin/airoboros-3.2: Processed 54517, Removed 4192
- PJMixers/grimulkan_theory-of-mind-ShareGPT: Processed 533, Removed 6
- grimulkan/PIPPA-augmented-dedup: Processed 869, Removed 46
- grimulkan/LimaRP-augmented: Processed 790, Removed 14
- PJMixers/grimulkan_physical-reasoning-ShareGPT: Processed 895, Removed 4
- MinervaAI/Aesir-Preview: Processed 994, Removed 6
- Doctor-Shotgun/no-robots-sharegpt: Processed 9911, Removed 89
Deduplication Stats:
Starting row count: 85628, Final row count: 81960, Rows removed: 3668