Full weight models
Select models uploaded in safetensors format. Currently all are merges. Annotations here.
Text Generation • Updated • 84 • 3Note I liked both versions, so this was literally a compromise merge.
grimjim/kukulemon-7B
Text Generation • Updated • 78 • 9Note Successful merge of kunoichi models with a roleplay model.
grimjim/kukulemon-spiked-9B
Text Generation • Updated • 22 • 3Note Somewhat successful frankenmerge. No fine-tune healing applied.
grimjim/kukulemon-32K-7B
Text Generation • Updated • 16 • 1Note Brought in 32k context from Mistral v0.2.
grimjim/kunoichi-lemon-royale-7B
Text Generation • Updated • 84 • 2Note Attempt to improve model intelligence by adding a high benchmark model to the formula.
grimjim/kunoichi-lemon-royale-v2-32K-7B
Text Generation • Updated • 18 • 5Note Brought in 32k context from Mistral v0.2.
grimjim/Mistral-7B-Instruct-demi-merge-v0.2-7B
Text Generation • Updated • 70Note Blended base and instruct for Mistral v0.2 for use in future merges.
grimjim/fireblossom-32K-7B
Text Generation • Updated • 28 • 2Note Partial success at infusing more creativity.
grimjim/zephyr-beta-wizardLM-2-merge-7B
Text Generation • Updated • 8 • 1Note Merge attempt to harness the capabilities of the contributing models.
grimjim/Mistral-Starling-merge-trial1-7B
Text Generation • Updated • 73Note A test. Feel free to skip.
grimjim/Mistral-Starling-merge-trial3-7B
Text Generation • Updated • 71Note A more successful test than trial1.
grimjim/zephyr-wizard-kuno-royale-BF16-merge-7B
Text Generation • Updated • 7 • 1Note An attempt to improve model intelligence by merging in a higher benchmarked model.
grimjim/llama-3-experiment-v1-9B
Text Generation • Updated • 9 • 4Note Benchmarks were for the most part unaffected by the addition of early middle layers; no fine-tuning was needed to heal damage the model.
grimjim/madwind-wizard-7B
Text Generation • Updated • 11 • 4Note Model intelligence was slightly damaged by the merge, but creativity is high.
grimjim/llama-3-merge-pp-instruct-8B
Text Generation • Updated • 27Note Merged back into Instruct to see what would happen.
grimjim/llama-3-merge-virt-req-8B
Text Generation • Updated • 13 • 2Note A requested merge.
grimjim/kukuspice-7B
Text Generation • Updated • 8Note Experiment. Limited success.
grimjim/cuckoo-starling-32k-7B
Text Generation • Updated • 10 • 2Note Fairly creative model.
grimjim/rogue-enchantress-32k-7B
Text Generation • Updated • 13 • 8Note Highly creative model, albeit a bit damaged. Could probably be repaired with a bit of fine-tuning.
grimjim/Mistral-7B-Instruct-demi-merge-v0.3-7B
Text Generation • Updated • 7Note Merge of base and instruct versions of Mistral v0.3. Unremarkable, unfortunately.
grimjim/kunoichi-lemon-royale-v3-32K-7B
Text Generation • Updated • 8 • 4Note Creativity was enhanced at some cost to format adherence.
grimjim/koboldai-erebus-extended-32k-7B
Text Generation • Updated • 7 • 3Note Attempt to merge in 32k context from Mistral v0.2.
grimjim/koboldai-holodeck-extended-32k-7B
Text Generation • Updated • 18 • 2Note Attempt to merge in 32k context from Mistral v0.2.
grimjim/kukulemon-v3-soul_mix-32k-7B
Text Generation • Updated • 20 • 4Note Enhanced creativity with only minor harm to the ability to follow format.
grimjim/Llama-3-Luminurse-v0.1-OAS-8B
Text Generation • Updated • 10 • 3Note The contribution of the biomedical model appears to have improved the intelligence of the resulting merge.
grimjim/Llama-3-Oasis-v1-OAS-8B
Text Generation • Updated • 16 • 4Note A merger of 3 models, all of which were effectively decensored by having refusal pathways ablated away.
grimjim/Llama-3-Steerpike-v1-OAS-8B
Text Generation • Updated • 14 • 1Note Imperfect model, but interesting.
grimjim/Llama-3-Luminurse-v0.2-OAS-8B
Text Generation • Updated • 255 • 6Note Further refinement of the Luminurse merge formula by incorporating another model specifically trained for consistency in following format when generating output.
grimjim/llama-3-Nephilim-v1-8B
Text Generation • Updated • 37 • 1Note The contribution of an offensive/defensive cybersecurity model (does that make it cyberwarfare model?) dramatically increased the creativity of the base model despite the low contributing weight.
grimjim/llama-3-Nephilim-v2-8B
Text Generation • Updated • 24 • 1Note Added contributions from two SOTA preference optimization models as well as a highly-trained roleplay model and a biomedical model.
grimjim/Llama-3-Instruct-8B-SPPO-Iter3-SimPO-merge
Text Generation • Updated • 2.44k • 2Note Merge of two PO models.
grimjim/Llama-3-Instruct-8B-SimPO-SPPO-Iter3-merge
Text Generation • Updated • 2.44k • 2Note Merge of two PO models.
grimjim/llama-3-Nephilim-v2.1-8B
Text Generation • Updated • 14 • 1Note A modest variation from v2.
grimjim/llama-3-Nephilim-v3-8B
Text Generation • Updated • 533 • 11Note This merge is one of the most coherent that I've produced.
grimjim/Llama-3.1-8B-Instruct-abliterated_via_adapter
Text Generation • Updated • 3.82k • 26Note It turns out a LoRA extracted from Llama 3 can successfully be applied to Llama 3.1. It works well, apparently.
grimjim/Llama-Nephilim-Metamorphosis-v1-8B
Text Generation • Updated • 23 • 1Note Proof that a Llama 3 model can be merged onto a Llama 3.1 using low weighting. It's not the best model as is, but is interesting due to its implications regarding what can be accomplished via merger.
grimjim/Magot-v1-Gemma2-8k-9B
Text Generation • Updated • 18 • 2grimjim/Llama-3.1-SuperNova-Lite-lorabilterated-8B
Text Generation • Updated • 34 • 3