Baselines

[rbroc]

On top of the feature-based classifiers, we want to have the following baselines:

• Tf-Idf (already implemented)
• Sentence embedding-based
• LLM-based classifier (e.g., GPT)

[MinaAlmasi]

Sentence embedding-based (sentence transformers)

[MinaAlmasi]

Sentence Embedding-based -> Sentence Transformers

Currently have embedded with a somewhat big Nvidia-model (see #80 desc.), but we can scale down e.g., with an e5 model.

Notes for that:

1. Need to check up again, but pretty sure the e5 models require a "passage: " prompt in front of all rows with text, which the current Nvidia model does not.
2. We might also not need to scale down to FP16 (with model.half()) if we use a smaller model,

[MinaAlmasi]

For LLM detector (bookkeeping)

• Want to use a "small" (3b-ish max) model that we we will not be using for generation
• Ruling out phi-3.5-mini since it has been trained on a lot of synthetic data