toddnief bidirectional-reversal issues

toddnief / bidirectional-reversal

0 stars 0 forks source link

issues

Newest

Newest Most commented Recently updated Oldest Least commented Least recently updated

Convert the misc test set to the QA and rephrased format

#41 toddnief opened 1 day ago
0
There's a problem with the token splitting for the generation eval

#40 toddnief opened 1 week ago
0
Localize behavior to K,Q,V?

#39 toddnief opened 1 week ago
0
Check entropy of wikitext (high accuracy but high loss)

#38 toddnief opened 1 week ago
0
Does finetuning on any data cause the same forgetting issue? How localized is the forgetting?

#37 toddnief closed 1 day ago
0
What if we have, say, Matt Damon in both positions? Does that help?

#36 toddnief opened 2 weeks ago
0
What if we have the "reversed "entity appear first but in a different kind of "slot"?

#35 toddnief opened 2 weeks ago
0
Create pipeline so that we can evaluate 60 examples (20 of each type)

#34 toddnief closed 1 week ago
1
Try finetuning a larger model using Cohere (maybe LoRA or not?)

#33 toddnief opened 2 weeks ago
0
Idea for testing in a non-synthetic setting: Evaluate full documents (perplexity should go way up from initial), replace all names with names from your fine-tuning set (perplexity should go back down)

#32 toddnief opened 2 weeks ago
0
Get the cumulative perplexity of each of the possible names in the training set as an eval

#31 toddnief closed 2 weeks ago
0
Do 1000 generations with just <BOS> from the final checkpoint of one of these models

#30 toddnief opened 1 month ago
0
Make sure huggingface trainer is shuffling the dataset

#29 toddnief closed 2 weeks ago
0
Figure out how to get the per token perplexity with colors for eval

#28 toddnief closed 1 month ago
0
Use spacy to parse names in eval sets

#27 toddnief closed 2 weeks ago
0
Add another language to eval set

#26 toddnief opened 1 month ago
0
Maybe do a generation on eval also? (just to see what's going on)

#25 toddnief closed 2 weeks ago
0
Use log likelihood of specific tokens for validation (as well as log likelihood of entire completion)

#24 toddnief closed 1 month ago
0
Try some combination of real and fake entities (ie real movies with fake entities)

#23 toddnief opened 2 months ago
0
Does the actual name matter? ie "Tom" vs "Joaquin"?

#22 toddnief opened 2 months ago
0
What if we do a bunch of paraphrases in the forward direction, does that "fix" reversal?

#21 toddnief opened 2 months ago
0
Try selecting 10 names from The Pile — one at each order of magnitude

#20 toddnief opened 2 months ago
0
Ratio of reversed vs unreversed facts

#19 toddnief opened 4 months ago
1
Make sure the embeddings and unembeddings are actually updated with full model SFT

#18 toddnief opened 4 months ago
0
Broadly there's probably some issue with the unembedding » try training a tied embedding/unembedding model?

#17 toddnief closed 2 months ago
0
What if we do SFT with no positional embedding...then do it with positional embedding

#16 toddnief opened 4 months ago
0
Add IMDB dataset to fine-tuning to add additional structured data so we don't get weird degenerate completions

#15 toddnief closed 1 month ago
0
Train 1B Pythia model to memorize to make sure we're not screwing up

#14 toddnief closed 1 day ago
0
Train 7B model with LoRA

#13 toddnief opened 4 months ago
0
Try a larger model (1B params) — Pythia?

#12 toddnief closed 4 months ago
0
Next company example (simpler and larger dataset)

#11 toddnief closed 4 months ago
1
Do masked models learn relations weirdly in general? (Eamon's question)

#10 toddnief opened 5 months ago
0
What about ROME edits with the Reversal Curse? (somehow related ideas)

#9 toddnief opened 5 months ago
1
What happens with few shot prompting?

#8 toddnief closed 5 months ago
0
Train with positional embedding reversed

#7 toddnief opened 5 months ago
0
Try training with smaller batch size

#6 toddnief closed 1 day ago
0
Fix logging behavior

#5 toddnief closed 1 day ago
0
Larger datasets: scaling laws for reversal curse?

#4 toddnief opened 6 months ago
0
Create company dataset: "X is a small company. Y is an employee of X. Z is coworkers with Y."

#3 toddnief closed 5 months ago
0
Try teammates with sport changed in test set...what happens?

#2 toddnief closed 5 months ago
0
Try teammates with first token of name masked

#1 toddnief closed 5 months ago
0