https://colab.research.google.com/drive/1mhcgx2SU3GrDq3pMZp_-JPtE_fO-7kGg#scrollTo=7GpmMGjpioid

simple_analogies_circuits.ipynb :

small and large: number seq and simple analogy using in-context learning. Use diff ways to corrupt and format inputs. Test on logit diff, attn pats, actv patch. based on exploratory demo

GPT-2-small prompts:

• "1 2 3 4”
• "2 4”
• "Mary is a human. Fido is a dog. The pet of this family is Fido. Rachel is a human. Pebbles is a cat. The pet of this family is”
• "The human is Mary. The dog is Fido. The pet is Fido. The human is Rachel. The dog is Pebbles. The pet is”

GPT-2-large prompts:

• "Mary is a teacher. Fido is a student. The child is Fido. Pebbles is a student. Rachel is a teacher. The child is”
  • Gets the most recent subject instead of the student
  • TRY: more in-context examples
• "Mary. Fido. Fido. Pebbles. Rachel. “
  • Fails
• "Mary is X. Fido is Y. Z is Fido. Pebbles is Y. Rachel is X. Z is “
  • Fails

https://colab.research.google.com/drive/1aOEeY4roW8oWqkZ0MuuZRJXmJGDRNcbr

simple_analogies_circuits, pt2.ipynb

Checks if GPT-2-Large and GPT-2-xl and GPT-Neo-2.7b can correctly complete analogies. Does preliminary logit lens, activation patching and attention head analysis on a few input formats. Also checks why these billion parameter models have trouble finding who “is” or who “has” given a previous statement from in-context learning. Prompts include:

• Transitivity: Mary has X. John has Y. Z is John. [Repeat template to find Z in new target system]
• "John is big. Mary is small. John is tall. Mary is”

Try:

• Use more in-context examples