Seagull Large Language Model
About:
- Re-implemented "Seagull," a transformer-based language model, as a part of Homework 4 in Cornell's CS 4740 Natural Language Processing Course course
- Fine-tuned model to generate humorous captions from descriptions of cartoons, leveraging tokenization, padding, multi-head attention, and feed-forward neural networks
- Streamlined model training with CUDA for faster iterations and outperformed humor benchmarks on large datasets
- Used advanced NLP techniques for context-aware captioning, achieving high relevancy and creativity in output
Tools and Technologies:
- PyTorch
- Google Colab
- Google Cloud GPU + TPU
Example Output
Input:
<|scene| data-preserve-html-node="true"> A politician is standing at a podium and delivering a speech in the middle of the forest. He has two secret service members guarding him. In the forest there are a bunch of animals hiding behind trees while watching the speech. <|uncanny| data-preserve-html-node="true"> The politician is giving a speech to animals rather than humans. >>>
**Model Output: **
I'm afraid this is what happens when the budget gets a little too far.
Input:
<|scene| data-preserve-html-node="true"> A man is looking out the window. There is a big statue right outside. <|uncanny| data-preserve-html-node="true"> The Easter Island statue is in this guy's yard. <|caption! data-preserve-html-node="true">
Model Output:
...and the next time I see it, I'll be in the shower
Technical Architecture
- Image 1: Full Seagull Transformer-based large language model. **Source: **https://github.coecis.cornell.edu/cs4740-fa23-public/hw4-fa23/blob/main/notebooks/hw4.ipynb
- Image 2: Token (and positional) embeddings **Source: **https://github.coecis.cornell.edu/cs4740-fa23-public/hw4-fa23/blob/main/notebooks/hw4.ipynb
- Image 3: Multi-headed Attention **Source: **https://github.coecis.cornell.edu/cs4740-fa23-public/hw4-fa23/blob/main/notebooks/hw4.ipynb
Citation:
- Authors: Tushaar Gangavarapu, Darren Key🦙, Logan Kraver🦙, Lionel Tan🦙, Pun Chaixanien🦜, Kai Horstmann🦜, Dave Jung🦜, Aaishi Uppuluri🦜. 2023. [CS 4740 Fa'23 HW4] Hush, the seagulls are purring: On generating humorous captions from scene descriptions.
- GitHub: https://github.coecis.cornell.edu/cs4740-fa23-public/hw4-fa23/.
