Causal program-aided language (CPAL) chain
The CPAL chain builds on the recent PAL to stop LLM hallucination. The problem with the PAL approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination.
The original PR's description contains a full overview.
Using the CPAL chain, the LLM translated this
"Tim buys the same number of pets as Cindy and Boris."
"Cindy buys the same number of pets as Bill plus Bob."
"Boris buys the same number of pets as Ben plus Beth."
"Bill buys the same number of pets as Obama."
"Bob buys the same number of pets as Obama."
"Ben buys the same number of pets as Obama."
"Beth buys the same number of pets as Obama."
"If Obama buys one pet, how many pets total does everyone buy?"
into this
.
Outline of code examples demoed in this notebook.
- CPAL's value against hallucination: CPAL vs PAL
1.1 Complex narrative
1.2 Unanswerable math word problem - CPAL's three types of causal diagrams (The Book of Why).
2.1 Mediator
2.2 Collider
2.3 Confounder
from IPython.display import SVG
from langchain.experimental.cpal.base import CPALChain
from langchain.chains import PALChain
from langchain import OpenAI
llm = OpenAI(temperature=0, max_tokens=512)
cpal_chain = CPALChain.from_univariate_prompt(llm=llm, verbose=True)
pal_chain = PALChain.from_math_prompt(llm=llm, verbose=True)
API Reference:
- PALChain from
langchain.chains
CPAL's value against hallucination: CPAL vs PAL
Like PAL, CPAL intends to reduce large language model (LLM) hallucination.
The CPAL chain is different from the PAL chain for a couple of reasons.
CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities.
1.1 Complex narrative
Takeaway: PAL hallucinates, CPAL does not hallucinate.
question = (
"Tim buys the same number of pets as Cindy and Boris."
"Cindy buys the same number of pets as Bill plus Bob."
"Boris buys the same number of pets as Ben plus Beth."
"Bill buys the same number of pets as Obama."
"Bob buys the same number of pets as Obama."
"Ben buys the same number of pets as Obama."
"Beth buys the same number of pets as Obama."
"If Obama buys one pet, how many pets total does everyone buy?"
)
pal_chain.run(question)
> Entering new chain...
def solution():
"""Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?"""
obama_pets = 1
tim_pets = obama_pets
cindy_pets = obama_pets + obama_pets
boris_pets = obama_pets + obama_pets
total_pets = tim_pets + cindy_pets + boris_pets
result = total_pets
return result
> Finished chain.
'5'
cpal_chain.run(question)
> Entering new chain...
story outcome data
name code value depends_on
0 obama pass 1.0 []
1 bill bill.value = obama.value 1.0 [obama]
2 bob bob.value = obama.value 1.0 [obama]
3 ben ben.value = obama.value 1.0 [obama]
4 beth beth.value = obama.value 1.0 [obama]
5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob]
6 boris boris.value = ben.value + beth.value 2.0 [ben, beth]
7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris]
query data
{
"question": "how many pets total does everyone buy?",
"expression": "SELECT SUM(value) FROM df",
"llm_error_msg": ""
}
> Finished chain.
13.0
# wait 20 secs to see display
cpal_chain.draw(path="web.svg")
SVG("web.svg")
Unanswerable math
Takeaway: PAL hallucinates, where CPAL, rather than hallucinate, answers with "unanswerable, narrative question and plot are incoherent"
question = (
"Jan has three times the number of pets as Marcia."
"Marcia has two more pets than Cindy."
"If Cindy has ten pets, how many pets does Barak have?"
)
pal_chain.run(question)
> Entering new chain...
def solution():
"""Jan has three times the number of pets as Marcia.Marcia has two more pets than Cindy.If Cindy has ten pets, how many pets does Barak have?"""
cindy_pets = 10
marcia_pets = cindy_pets + 2
jan_pets = marcia_pets * 3
result = jan_pets
return result
> Finished chain.
'36'
try:
cpal_chain.run(question)
except Exception as e_msg:
print(e_msg)
> Entering new chain...
story outcome data
name code value depends_on
0 cindy pass 10.0 []
1 marcia marcia.value = cindy.value + 2 12.0 [cindy]
2 jan jan.value = marcia.value * 3 36.0 [marcia]
query data
{
"question": "how many pets does barak have?",
"expression": "SELECT name, value FROM df WHERE name = 'barak'",
"llm_error_msg": ""
}
unanswerable, query and outcome are incoherent
outcome:
name code value depends_on
0 cindy pass 10.0 []
1 marcia marcia.value = cindy.value + 2 12.0 [cindy]
2 jan jan.value = marcia.value * 3 36.0 [marcia]
query:
{'question': 'how many pets does barak have?', 'expression': "SELECT name, value FROM df WHERE name = 'barak'", 'llm_error_msg': ''}
Basic math
Causal mediator
question = (
"Jan has three times the number of pets as Marcia. "
"Marcia has two more pets than Cindy. "
"If Cindy has four pets, how many total pets do the three have?"
)
PAL
pal_chain.run(question)
> Entering new chain...
def solution():
"""Jan has three times the number of pets as Marcia. Marcia has two more pets than Cindy. If Cindy has four pets, how many total pets do the three have?"""
cindy_pets = 4
marcia_pets = cindy_pets + 2
jan_pets = marcia_pets * 3
total_pets = cindy_pets + marcia_pets + jan_pets
result = total_pets
return result
> Finished chain.
'28'
CPAL
cpal_chain.run(question)
> Entering new chain...
story outcome data
name code value depends_on
0 cindy pass 4.0 []
1 marcia marcia.value = cindy.value + 2 6.0 [cindy]
2 jan jan.value = marcia.value * 3 18.0 [marcia]
query data
{
"question": "how many total pets do the three have?",
"expression": "SELECT SUM(value) FROM df",
"llm_error_msg": ""
}
> Finished chain.
28.0
# wait 20 secs to see display
cpal_chain.draw(path="web.svg")
SVG("web.svg")
Causal collider
question = (
"Jan has the number of pets as Marcia plus the number of pets as Cindy. "
"Marcia has no pets. "
"If Cindy has four pets, how many total pets do the three have?"
)
cpal_chain.run(question)
> Entering new chain...
story outcome data
name code value depends_on
0 marcia pass 0.0 []
1 cindy pass 4.0 []
2 jan jan.value = marcia.value + cindy.value 4.0 [marcia, cindy]
query data
{
"question": "how many total pets do the three have?",
"expression": "SELECT SUM(value) FROM df",
"llm_error_msg": ""
}
> Finished chain.
8.0
# wait 20 secs to see display
cpal_chain.draw(path="web.svg")
SVG("web.svg")
Causal confounder
question = (
"Jan has the number of pets as Marcia plus the number of pets as Cindy. "
"Marcia has two more pets than Cindy. "
"If Cindy has four pets, how many total pets do the three have?"
)
cpal_chain.run(question)
> Entering new chain...
story outcome data
name code value depends_on
0 cindy pass 4.0 []
1 marcia marcia.value = cindy.value + 2 6.0 [cindy]
2 jan jan.value = cindy.value + marcia.value 10.0 [cindy, marcia]
query data
{
"question": "how many total pets do the three have?",
"expression": "SELECT SUM(value) FROM df",
"llm_error_msg": ""
}
> Finished chain.
20.0
# wait 20 secs to see display
cpal_chain.draw(path="web.svg")
SVG("web.svg")
%autoreload 2