Pietro Lesci

pietrolesci

https://pietrolesci.github.io/

AI & ML interests

I like developing and applying causal methods to study the effect of training choices on models’ behaviour, including memorisation, shortcut learning, and tokenisation.

Recent Activity

updated a dataset 7 days ago

pietrolesci/me-minipile-evals

updated a collection 7 days ago

Merge-Effect

updated a dataset 23 days ago

pietrolesci/pile-deduped

View all activity

Organizations

pietrolesci's activity

New activity in JeanKaddour/minipile 8 months ago

Domain and provenance annotation

#1 opened over 1 year ago by

haukur

New activity in HuggingFaceTB/SmolLM-135M 9 months ago

Trapezoidal scheduler with cooldown phase

👍 1

#4 opened 11 months ago by

maveriq

New activity in EleutherAI/pythia-160m about 1 year ago

Tokenizer `merges.txt` files

#5 opened about 1 year ago by

pietrolesci

New activity in EleutherAI/pile-deduped-pythia-preshuffled over 1 year ago

Sequence "packing" logic

👍 2

#2 opened over 1 year ago by

pietrolesci

Pad-only sequences from mmap'ed dataset after a certain index

#1 opened over 1 year ago by

pietrolesci

New activity in EleutherAI/pile-duped-pythia-random-sampled over 1 year ago

Add full sequences (beyond the first 64 tokens)

#1 opened over 1 year ago by

pietrolesci

Add full sequences (beyond the first 64 tokens)

#1 opened over 1 year ago by

pietrolesci

New activity in JeanKaddour/minipile over 1 year ago

Domain and provenance annotation

#1 opened over 1 year ago by

haukur