@anamarasovic@sigmoid.social

Ana Marasović @anamarasovic@sigmoid.social

· 46mo ago

Everyone is over this, but I'm still amazed that we have in-context learning, it's even a solid baseline #nlproc #nlp

4

0

Ana Marasović @anamarasovic@sigmoid.social

· 46mo ago

Y'all my dad got me a personalized jersey 🥹

4

0

Ana Marasović @anamarasovic@sigmoid.social

· 46mo ago

1. Don't accept reviewing without carefully checking your calendar

2. Don't be late with your reviews without notifying ACs on time

3. Don't ignore ACs

4. Don't be late again to a new set date

5. Don't disappear when the discussion starts

#ACL2023

1/

3

0

Ana Marasović @anamarasovic@sigmoid.social

· 46mo ago

Those who have resources, here is a request for model pretraining (😅):

Train decoder and encoder-decoder transformers with the same data/resources and some of the HPs (e.g. # layers), with:

(1) the same objective e.g. LM

(2) objective(s) used in GPT, BART, T5

Many papers get different insights with GPT/BART/T5 (e.g. https://arxiv.org/abs/2210.13575) but no way of saying what causes these differences which leaves reviewers and readers (rightfully) displeased with observations

#NLProc #nlp

arXiv.org

Does Self-Rationalization Improve Robustness to Spurious Correlations?

Rationalization is fundamental to human reasoning and learning. NLP models trained to produce rationales along with predictions, called self-rationalization models, have been investigated for their interpretability and utility to end-users. However, the extent to which training with human-written rationales facilitates learning remains an under-explored question. We ask whether training models to self-rationalize can aid in their learning to solve tasks for the right reasons. Specifically, we ev

3

0

1

0

Ana Marasović @anamarasovic@sigmoid.social

· 46mo ago

Wow this got complicated so suddenly 👀 (🇫🇷🇦🇷)

2

0

Ana Marasović @anamarasovic@sigmoid.social

· 47mo ago

Used the edit function for the first time 🤤

2

0

1

0

Ana Marasović @anamarasovic@sigmoid.social

· 46mo ago

@soldni@mastodon.social hope it heals soon!

1

0

Ana Marasović @anamarasovic@sigmoid.social

· 46mo ago

Comment on this: https://twitter.com/_jasonwei/status/1602714749531312129

Responses say that the benchmarks are flawed as code-davinci-2 > text-davinci-2(3).

Why it's not speculated that OpenAI’s FT/RLFH/PPO and/or human demonstrations in the playground/finetuning data are faulty?

The results show that FLAN-PaLM > PaLM. Isn't that a better signal that instruction finetuning does NOT hurt performance on academic benchmarks given that PaLM is much better documented.

#NLProc #nlp

X (formerly Twitter)

Jason Wei (@_jasonwei) on X

- code-davinci-2 > text-davinci-3 means that their instruction finetuning overall hurts performance on academic benchmarks - text-davinci-3 > text-davinci-2 means that PPO improves performance https://t.co/XfJE1DWvtQ

1

0

1

0

Ana Marasović @anamarasovic@sigmoid.social

· 47mo ago

@soldni@mastodon.social ✔️

1

0

Ana Marasović @anamarasovic@sigmoid.social

· 44mo ago

How do you include inexperienced undergrads in NLP research? How were you included as an undergrad?

Talking about students who haven't yet completed all of the relevant coursework (to me DL, ML, NLP), or any research project.

There are many such students and giving them all even a tiny research project is not feasible because I'd need to spend a considerable amount of time going over core concepts typically covered in classes. 1/

0

1

0