TWIL ~ This week I learned"

Main message of this week, is that OpenCV is not a beginner friendly framework. So my pursuits in computer vision will take a lot longer than expected. Just to start…

langchain & co

I am finally going into LLM ecosystems available for commoners. Langchain seems dominating the space with their toolbox:

• langchain: core framework for connecting data, agents, llms and prompts
• streamlit-agent: example streamlit GPT access (demo: https://langchain-streaming-example.streamlit.app/)
• langsmith: cloud logging app, not open sourced
• langserve: API and setup scripts for langchain based apps

OpenSource RAG

I am on a small hunt for open source libs, that do RAGs well. Found two promising and going deeper on them:

• Neum.AI
• LlamaIndex
• OpenGPTs

Neum AI

First one I stumbled upon. I really like the connectors that also contains SharePoint. From documentation the Neum.ai is leaning more towards cloud deployment. But local options still seems ok.

https://www.neum.ai/

LlamaIndex

Llama seems the most advanced and elaborate as of now. Loading is not as Neum.ai, but the rest seems much more “there”. Storing, Indexing, Prompting is so varied, that even choosing starting template will take some time. Documentation seems clear enough, though.

https://gpt-index.readthedocs.io/en/latest/use_cases/q_and_a.html

OpenGPTs

Also langchain group come with their variation (preview https://opengpts-example-vz4y4ooboq-uc.a.run.app/). It is more of an “How to connect our tools” than someting new.

https://github.com/langchain-ai/opengpts/

Orca 2

https://arxiv.org/pdf/2311.11045.pdf

Bullet points from reading:

• (It seems) Almost every small LM is starting with Llama2 weights
• You can try Llama 2 13b here: https://replicate.com/meta/llama-2-13b?input=form&output=preview
• There is an open arena to test LLMs: https://chat.lmsys.org/
• Direct citation of resources used for training: Compute: We trained Orca 2 on 32 NVIDIA A100 GPUs with 80GB memory with bfloat16. For the 13B checkpoint, it took ~17 hours to train Orca 2 on FLAN dataset for one epoch, ~40 hours to train on 5 million ChatGPT data for 3 epochs and ~23 hours to continue training on ~1.8 million GPT-4 data for 4 epochs.
• Paper introduced Cautious Reasoning and also used syntentic data generated by GPT-4
• Research seems on the training data part and prompting/interactions between components during training. Wow.

Data Science Bulletin

After a long break, my dear friend published another issue of DSB: https://datasciencebulletin.com/dsbenglish/dsb-143/. Great read as always.