Update README.md

README.md

@@ -5,10 +5,9 @@ datasets:
 - trollek/Mouse-Diffusion-Instruct
 - trollek/CodeMouse
 - trollek/Panoia-v01
-- WhiteRabbitNeo/WRN-Chapter-1
-- WhiteRabbitNeo/WRN-Chapter-2
 - jondurbin/airoboros-3.2
 - mlabonne/orpo-dpo-mix-40k
+- Magpie-Align/Magpie-Air-300K-Filtered
 language:
 - en
 ---
@@ -17,8 +16,6 @@ language:
 
 A brand spanking new model with a silly name. Brought to you by Anoia, the Goddess of Things That Get Stuck in Drawers, and the psychologial damage of having optic nerves.
 
-It seems like the more LoRAs we cake on a base model, the more they hallucinate. (source to be added). So in the last installment of NinjaMouse2 I present to you, a delightful human being, a finetuned danube2 model that have been extended without the Llama Pro method of interleaving new layers/transformer blocks. Just regular ol' [mergekit](https://github.com/arcee-ai/mergekit).
-
 ```yaml
 slices:
   - sources:
@@ -32,3 +29,18 @@ dtype: bfloat16
 ```
 
 I tried several other layer configerations but this one had the least negative effects on Hellaswag and Winogrande evals. Next up I made different finetunes of danube2 using Badam and merged them with the Model Stock method.
+
+It uses the default template of danube2:
+
+```jinja2
+<|prompt|>{{instruction}}</s><|answer|>{{response}}</s>
+```
+
+And can be used with the Ollama ComfyUI extension:
+
+[<img src="https://huggingface.co/trollek/NinjaMouse2-2.5B-v0.2/resolve/main/ollama_comfyui.png" width="800px"/>](https://huggingface.co/trollek/NinjaMouse2-2.5B-v0.2/resolve/main/ollama_comfyui.png)
+
+[<img src="https://huggingface.co/trollek/NinjaMouse2-2.5B-v0.2/resolve/main/tophat_cat.png" width="800px"/>](https://huggingface.co/trollek/NinjaMouse2-2.5B-v0.2/resolve/main/tophat_cat.png)
+
+
+Trying to fine-tune the chat model even further was a mistake. This is the last mouse ninja.