grimjim/SauerHuatuoSkywork-o1-Llama-3.1-8B

Text Generation • Updated 15 days ago • 58 • 2

Note The addition of o1-inspired reasoning uplifted the Instruct model on most benchmarks. As of the initial merge release date, this is the second highest benching Llama 3.x 8B model that I've achieved on the newer Open LLM leaderboard.

grimjim/DeepSauerHuatuoSkywork-R1-o1-Llama-3.1-8B

Text Generation • Updated 4 days ago • 41 • 4

Note Merging in a touch of DeepSeek R1 distillation improved benchmarks more than it hurt them. This is currently my highest benching Llama 3.x 8B model on the newer Open LLM Leaderboard.

grimjim/HuatuoSkywork-o1-Llama-3.1-8B

Text Generation • Updated 28 days ago • 112

Note This merge of o1 reasoning models achieved an unexpectedly high MATH Level 5 score of 33.99%, which was the highest I saw at the time for Llama 3.x 8B models on the Open LLM Leaderboard.

grimjim/llama-3-Nephilim-v3-8B

Text Generation • Updated Sep 3, 2024 • 135 • 13

Note Proof of concept that a text completion model, based on Instruct in this case, doesn't need any fine-tuning specifically targeting roleplay. All merge components are academic in origin.

grimjim/Llama-3.1-8B-Instruct-abliterated_via_adapter

Text Generation • Updated Sep 18, 2024 • 5.31k • 29

Note Llama 3.1 8B "abliterated" via transfer of the feature via a LoRA. There's probably some damage to the model that could be fixed with additional fine-tuning, as that's a common consequence of abliteration.

grimjim/Llama-3-Instruct-abliteration-LoRA-8B

Updated Sep 10, 2024 • 7

Note The LoRA adapter obtained from Llama 3, and later applied against Llama 3.1.

grimjim/kukulemon-7B

Text Generation • Updated Mar 21, 2024 • 52 • 11

Note One of my first merges, combining two smart models with a roleplay-oriented merge. Someone on YouTube called out this Mistral v0.1 7B architecture model in a video.