Skip to content

Latest commit

 

History

History
83 lines (57 loc) · 3.77 KB

README.md

File metadata and controls

83 lines (57 loc) · 3.77 KB
Raw

yeLLowhaMmer

Working repository for Team datalab in the 2024 LLM Hackathon for Applications in Materials & Chemistry.

📹 Final demo video (link)

Plan

Create an LLM-based agent that can interact with the datalab Python API so we don't have to implement things ourselves (as developers of datalab). This follows on from our work on Whinchat 🐦 last year, where we integrated GPT-based models from OpenAI with datalab such that users can interact conversationally with their raw data (see Jablonka et al., 10.1039/D3DD00113J).

Potential applications:

  • Using datalab to parse datafiles, return e.g., dataframes then using the LLM to create plots of data we don't support as blocks, like comparing XRD data from multiple samples
  • Upload a picture of a lab notebook page, then use the LLM to read it and make the appropriate API call to add it to datalab
  • Many more! (pls add them)

Log of experiments

  • playground/summarise_public_datalab.py Using shroominic/codeinterpreter-api to generate Python scripts that are executed locally that do the prompted task.
    • Works up to a point (at least with latest models, e.g., Claude 3 Opus [expensive]) but requires lots of back-and-forth to generate valid code.
    • Most of the problems are related to either
      1. eccentricities of our API (e.g., needing to know ahead of time magic strings for item types and otherwise)
      2. weird behaviour where syntactically invalid code is created for the first few iterations
    • Therefore, at least some time during this hackathon will be made making our API package more bot- (and hopefully human-)friendly, via more examples and more ergonomic tweaks (e.g., automatically pluralising item types if the singular is given).
  • streamlit_app/app.py uses a hacked version of CodeBox to provide a chat UI and run generated code.
    • We hacked it so that images/plots can be returned, as well as being able to copy the generate code for further human tweaking.
    • With Anthropic models, it keeps trying to generate and execute code until it works.
    • The session is persistent, so you can then ask it to do things with the variables it creates, although often it tries to start again from scratch anyway.

Setup

Install dependencies

Make a Python environment (with whatever method you prefer) and install the deps from the requirements.txt file, e.g., using the standard library:

python -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt

or using something like uv:

uv venv
uv pip install -r requirements.txt
. .venv/bin/activate

If you want to add a dependency, you can add it there. Occasionally we will also generate a lockfile to make sure all have compatible requirements.

Generate lock:

uv pip compile --prerelease=allow -r requirements.txt > requirements-lock.txt

Install lock (uv):

uv pip install ---prerelease=allow -r requirements-lock.txt

or directly in your virtual environment:

pip install -r requirements-lock.txt

Running the Streamlit app

After dependencies have been installed, you can launch the Streamlit app with:

streamlit run streamlit_app/app.py