Priority Task : start using trulens to evaluate Gemini

[Josephrp]

🤔How To

Check our References

• sample app : https://huggingface.co/spaces/Tonic/Gemini/
• gemini api : https://console.cloud.google.com/vertex-ai/generative/multimodal/create/text?project=focused-album-408018
• gemini docs : https://cloud.google.com/vertex-ai/docs/generative-ai/model-reference/gemini?_ga=2.179001954.-799536016.1700120910&_gac=1.26697039.1702492576.Cj0KCQiAyeWrBhDDARIsAGP1mWRBfEyPIju5t-GUBCU4kfiqJq_YQAYEbuQE9qwfgi82tBqdYiiorp4aAuD0EALw_wcB

trulens github + notebooks : https://github.com/truera/trulens/tree/main/trulens_eval/examples

• RAG Evaluation : https://lablab.ai/t/trulens-google-vertex-ai-tutorial-building-rag-applications
• Comparing models with TruLens : https://github.com/truera/trulens/blob/main/trulens_eval/examples/expositional/use_cases/model_comparison.ipynb
• Llama Index / Chain of Reasoning Eval with Trulens :
  https://github.com/truera/trulens/blob/main/trulens_eval/examples/expositional/frameworks/llama_index/llama_index_complex_evals.ipynb
• Llama Index Retrieval Quality : https://github.com/truera/trulens/blob/main/trulens_eval/examples/expositional/frameworks/llama_index/llama_index_retrievalquality.ipynb

Ideas for Evaluation

• RAG
• System Prompt
• Data Processing Pipeline
• Image Inputs

Work

What it takes : literally just running a notebook.

• chroma , or embeddings to test
• list of prompts to test
• test combinations of prompts
• multimodal evaluations

we will include the notebooks in the submission and write up

[Josephrp]

hey there @mie-h and @Zochory : https://github.com/Tonic-AI/DataTonic/tree/main/evaluation this is a folder where we will first start working on the trulens evaluations which are a hackathon requirement + good practice while building an app 🫡

[Josephrp]

hey there @mie-h & @Zochory : i added default prompts to the baseline prompts folder we can use those in a trulens evaluation.

[Josephrp]

consider using this to generate "system prompts" for gemini

[Josephrp]

added an incomplete example notebook : https://github.com/Tonic-AI/DataTonic/blob/main/evaluation/results/modelcomparision.ipynb

[Josephrp]

big thank you to 🏆😎 @MN-Noor for producing the first TruLens with gemini on RAG using open ai!

Open tasks :

• Make a notebook to test Gemini MultiModal (image inputs)
• Make a notebook to test more models against Gemini
• Make a notebook to test the "new features of Gemini" like the censorship level.

we'll all work on this together, normally if everyone does one, or at least contributes to a good one we will have secured this task.

[Zochory]

Est-ce que l'on ajouterait pas d'autres multimodal LLM ?
comme celui ci dans les evals ? https://huggingface.co/sshh12/Mistral-7B-LoRA-ImageBind-LLAVA