shisa-ai/shisa-v1-llama3-70b

Text Generation • Updated 24 days ago • 66 • 2

Note 2024-05: The shisa-v1 dataset applied to Llama 3 Instruct 70B outperforms gpt-3.5-turbo

shisa-ai/shisa-v1-llama3-8b

Text Generation • Updated 24 days ago • 124 • 3

Note 2024-05: The shisa-v1 dataset applied to Llama 3 Instruct 8B leads to significantly improved performance

augmxnt/shisa-gamma-7b-v1

Text Generation • Updated May 19 • 17.8k • 13

Note 2023-12: A version using the shisa-v1 dataset applied to Japanese Stable LM Base Gamma 7B. Less tokenizer efficiency, but better overall performance

augmxnt/shisa-7b-v1

Text Generation • Updated Dec 20, 2023 • 2.35k • 28

Note 2023-12: In addition to SFT, this also underwent a DPO round which improved human preference rating

Note Largely synthetic dataset combining Airoboros, Ultrachat, Orca in JA and EN. Also, the Jaster train set

augmxnt/shisa-base-7b-v1

Text Generation • Updated Dec 9, 2023 • 2.39k • 15

Note 2023-12: A continued pre-train (8B 90% JA tokens) of Mistral 7B v0.1 w/ tokenizer extension; probably needs 10B more tokens of pretraining tbt