Week 2

Sept 5th 2023

• Checked the VS_Research folder on WD_Black. Using rclone to compare sizes on GDrive and local as following:

  • GDrive:

      (base) gqc@msi-desktop-ubuntu:WD_BLACK1$ rclone size hydrotrek-gdrive:/VS_Research
            2023/09/05 09:54:25 NOTICE: Dangling shortcut "2022 Metro Van CCTV Data" detected
            2023/09/05 09:54:30 NOTICE: Dangling shortcut "January 2023 DNV upload" detected
            2023/09/05 10:32:50 NOTICE: Google drive root 'VS_Research': Size may be underestimated due to 15 objects with unknown size
            Total objects: 2.571M (2571131)
            Total size: 7.275 TiB (7999326983846 Byte)
            Total objects with unknown size: 15 (15)

  • WD_Black:

      (base) gqc@msi-desktop-ubuntu:WD_BLACK1$ du -hs ./VS_Research/
            2.6T    ./VS_Research/

• Working on using LLAMA playground here -> llama2.ai

• FYI, ChatGPT playground is here -> platform.openai.com

• Resources used for learning and testing implementation of LLAMA -

  • List of resources
  • maginative.com
  • hf.co/docs
  • hf.co/blog

Sept 6th 2023

• Research Topic for later - Combining output from two different models.

• Mail AI3 questions to - ai3info@lists.tamu.edu; Lizzette Arias lizzette.arias@us-ignite.org

• Create a LLAMA page under AI3 in projects site for adding details and marmaid charts.

  • Add notes on what are being used in the scripts.
  • Note the wandb token used for fine-tuning LLaMA.

• Included the LLAMA notebooks in AI3-hackathon/docs/nbs/ - Not yet committed, also uploaded on colab.

• Add comments in the LLAMA 2 notebooks.

• Test the trained llama/generate.json using a new notebook or subscriber.py.

  • Works with subscriber.py, need to test on more data.

• Check if LLaMA has an API (we know HF(hugging face) hosts it) - No offical API, only third party hosted APIs

• Should we train LLaMA on phase 3 data? - Yes

• Improve YOLO output (misclassification)

  • Sudhir mailed this to try - yolo

Sept 7th 2023

• Worked on running CCTV labelling notebook on MSI through GQC user - Since SSH session through VSCode seems to timeout.
• Add Jian's reply to questions on general site under AI3.
• Document LLAMA notebooks flowchart under projects > AI3 > LLAMA

Sept 8th 2023

• Decide on training LLAMA on new phase 3 data over the weekend.
  • Search on how to train a fine-tuned model with more data - will it forget the old data?
  • Prepare the lookup DB with synonyms and keys
  • Prepare the training DB with json, questions and answers[need to decide].
• Need to work on preparing lookup.db such that any new keys words are appended to existing list.
• Update data sources section on phase3 roadmap page.
• Create AI3 > videos section to discuss on our processes like YOLO.