Model Card for Model ID
Model Details
Model Description
Mind-wandering stream-of-consciousness type elaboration has been found useful in addressing wicked and ill-defined problems creatively, especially in the very first parts of the design process (problem redefinition). Mind-wandering benefits the creative process by highlighting multitude of ways to approach the design problem, without entering into concrete solutions. However, producing stream-of-consciousness type output is challenging for many people, especially in busy project life.
The purpose of this model is to use the hallucinative tendency of the LLMs as a benefit in the creative processes. The focus is in identifying potential design paradoxes (which, based on research, open doors for creative solutions). The model is fine-tuned to continue an input of "Here are my thoughts about the design paradox of [design problem here]" with a stream of consciousness -like text where chunks of freely-associating design paradox elaboration is folowed by quick jumps to next chunks. The result is detailed mind-wandering on the design context's design paradoxes.
Due to unstructured nature, the output of this model server only little purpose itself: therefore, the output can and should be systematically analyzed with more structured LLM's, such as OpenAI ChatGPT 4.0 (turbo). To identify design paradoxes and design directions, one can analyze the output, e.g., with this ChatGPT4.0 prompt: "From this text, go deep and create a list of unexpected design paradoxes that might stimulate creativity: Here are my thoughts about the design paradoxes of [design problem here]: [model output here]". After that, to systematically ideate on some identified paradox, one can use this ChatGPT prompt: "Connected to the problem of [design problem here], create unusual creative platform business ideas based on this design paradox (do not care if the idea is silly, if it is CREATIVE): [selected paradox from ChatGPT output]".
- Developed by: Ville Eloranta
- Funded by [optional]: n/a
- Shared by [optional]: n/a
- Model type: n/a
- Language(s) (NLP): n/a
- License: Apache 2.0
- Finetuned from model [optional]: Mistral-7b-v0.1 (non instruct model)
Model Sources [optional]
- Repository: villee/mistral01_streamofconsciousnessB_bat1lora8_gguf
- Paper [optional]: n/a
- Demo [optional]: n/a
Uses
The model is fine-tuned to continue an input of "Here are my thoughts about the design paradox of [design problem here]" with a stream of consciousness -like text where chunks of freely-associating design paradox elaboration is folowed by quick jumps to next chunks. The result is detailed mind-wandering on the design context's design paradoxes.
Due to unstructured nature, the output of this model server only little purpose itself: therefore, the output can and should be systematically analyzed with more structured LLM's, such as OpenAI ChatGPT 4.0 (turbo). To identify design paradoxes and design directions, one can analyze the output, e.g., with this ChatGPT4.0 prompt: "From this text, go deep and create a list of unexpected design paradoxes that might stimulate creativity: Here are my thoughts about the design paradoxes of [design problem here]: [model output here]". After that, to systematically ideate on some identified paradox, one can use this ChatGPT prompt: "Connected to the problem of [design problem here], create unusual creative platform business ideas based on this design paradox (do not care if the idea is silly, if it is CREATIVE): [selected paradox from ChatGPT output]".
Note:
- Usage: Prompt "Here are my thoughts about the design paradox of [design problem here]"
- the model should be used with high temperature (e.g., 0.8-1.0) and long contexts (e.g., 4192)
- by design the model might produce repetitive content - please break the repetition if the content no longer progresses
- model output might have weird format; that is also by design
Bias, Risks, and Limitations
There is no moderation in the model so use with own risk.
Recommendations
Not for production usage.
How to Get Started with the Model
- Start ollama in one shell: ollama serve
- In another shell, download the model: curl -L https://huggingface.co/villee/mistral01_streamofconsciousnessB_bat1lora8_gguf/resolve/main/streamofconsciousnessB_bat1lora8.gguf -o streamofconsciousnessB_bat1lora8.gguf
- Create Modelfile: FROM "streamofconsciousnessB_bat1lora8.gguf" PARAMETER temperature 1 PARAMETER num_ctx 4096
- Create ollama instance: ollama create streamofconsciousness -f Modelfile
- Infer with ollama: ollama run streamofconsciousness "Here are my thoughts about the design paradoxes of [design challenge]:"
- e.g. "Here are my thoughts about the design paradoxes of making the electricity markets more stable in a situation where the price of renewable power sources fluctuates wildly:"
- Rerun after you get a nice long stream of consciousness.
Training Details
Training Data
- the model is based on Mistral-7b-v0.1 (non instruct model)
- fine tune dataset is this: villee/streamofconsciousness (contains 200 rows of fine-tune data to enable Mistral to do stream-of-consciousness type output)
Training Procedure
- fine-tune has been done through lora (batch 1, lora layers 8) with apple mlx
