id
stringlengths 14
16
| text
stringlengths 29
2.73k
| source
stringlengths 49
115
|
|---|---|---|
c5c7551d58b9-0 | .ipynb
.pdf
Simulated Environment: Gymnasium
Contents
Define the agent
Initialize the simulated environment and agent
Main loop
Simulated Environment: Gymnasium#
For many applications of LLM agents, the environment is real (internet, database, REPL, etc). However, we can also define agents to interact in simulated environments like text-based games. This is an example of how to create a simple agent-environment interaction loop with Gymnasium (formerly OpenAI Gym).
!pip install gymnasium
Requirement already satisfied: gymnasium in /Users/michaelchang/.miniconda3/envs/langchain/lib/python3.9/site-packages (0.28.1)
Requirement already satisfied: farama-notifications>=0.0.1 in /Users/michaelchang/.miniconda3/envs/langchain/lib/python3.9/site-packages (from gymnasium) (0.0.4)
Requirement already satisfied: importlib-metadata>=4.8.0 in /Users/michaelchang/.miniconda3/envs/langchain/lib/python3.9/site-packages (from gymnasium) (6.0.1)
Requirement already satisfied: cloudpickle>=1.2.0 in /Users/michaelchang/.miniconda3/envs/langchain/lib/python3.9/site-packages (from gymnasium) (2.2.1)
Requirement already satisfied: numpy>=1.21.0 in /Users/michaelchang/.miniconda3/envs/langchain/lib/python3.9/site-packages (from gymnasium) (1.24.3)
Requirement already satisfied: jax-jumpy>=1.0.0 in /Users/michaelchang/.miniconda3/envs/langchain/lib/python3.9/site-packages (from gymnasium) (1.0.0) | https://python.langchain.com/en/latest/use_cases/agent_simulations/gymnasium.html |
c5c7551d58b9-1 | Requirement already satisfied: typing-extensions>=4.3.0 in /Users/michaelchang/.miniconda3/envs/langchain/lib/python3.9/site-packages (from gymnasium) (4.5.0)
Requirement already satisfied: zipp>=0.5 in /Users/michaelchang/.miniconda3/envs/langchain/lib/python3.9/site-packages (from importlib-metadata>=4.8.0->gymnasium) (3.15.0)
import gymnasium as gym
import inspect
import tenacity
from langchain.chat_models import ChatOpenAI
from langchain.schema import (
AIMessage,
HumanMessage,
SystemMessage,
BaseMessage,
)
from langchain.output_parsers import RegexParser
Define the agent#
class GymnasiumAgent():
@classmethod
def get_docs(cls, env):
return env.unwrapped.__doc__
def __init__(self, model, env):
self.model = model
self.env = env
self.docs = self.get_docs(env)
self.instructions = """
Your goal is to maximize your return, i.e. the sum of the rewards you receive.
I will give you an observation, reward, terminiation flag, truncation flag, and the return so far, formatted as:
Observation: <observation>
Reward: <reward>
Termination: <termination>
Truncation: <truncation>
Return: <sum_of_rewards>
You will respond with an action, formatted as:
Action: <action>
where you replace <action> with your actual action.
Do nothing else but return the action.
"""
self.action_parser = RegexParser(
regex=r"Action: (.*)",
output_keys=['action'], | https://python.langchain.com/en/latest/use_cases/agent_simulations/gymnasium.html |
c5c7551d58b9-2 | regex=r"Action: (.*)",
output_keys=['action'],
default_output_key='action')
self.message_history = []
self.ret = 0
def random_action(self):
action = self.env.action_space.sample()
return action
def reset(self):
self.message_history = [
SystemMessage(content=self.docs),
SystemMessage(content=self.instructions),
]
def observe(self, obs, rew=0, term=False, trunc=False, info=None):
self.ret += rew
obs_message = f"""
Observation: {obs}
Reward: {rew}
Termination: {term}
Truncation: {trunc}
Return: {self.ret}
"""
self.message_history.append(HumanMessage(content=obs_message))
return obs_message
def _act(self):
act_message = self.model(self.message_history)
self.message_history.append(act_message)
action = int(self.action_parser.parse(act_message.content)['action'])
return action
def act(self):
try:
for attempt in tenacity.Retrying(
stop=tenacity.stop_after_attempt(2),
wait=tenacity.wait_none(), # No waiting time between retries
retry=tenacity.retry_if_exception_type(ValueError),
before_sleep=lambda retry_state: print(f"ValueError occurred: {retry_state.outcome.exception()}, retrying..."),
):
with attempt:
action = self._act()
except tenacity.RetryError as e:
action = self.random_action()
return action
Initialize the simulated environment and agent#
env = gym.make("Blackjack-v1") | https://python.langchain.com/en/latest/use_cases/agent_simulations/gymnasium.html |
c5c7551d58b9-3 | Initialize the simulated environment and agent#
env = gym.make("Blackjack-v1")
agent = GymnasiumAgent(model=ChatOpenAI(temperature=0.2), env=env)
Main loop#
observation, info = env.reset()
agent.reset()
obs_message = agent.observe(observation)
print(obs_message)
while True:
action = agent.act()
observation, reward, termination, truncation, info = env.step(action)
obs_message = agent.observe(observation, reward, termination, truncation, info)
print(f'Action: {action}')
print(obs_message)
if termination or truncation:
print('break', termination, truncation)
break
env.close()
Observation: (15, 4, 0)
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 1
Observation: (25, 4, 0)
Reward: -1.0
Termination: True
Truncation: False
Return: -1.0
break True False
Contents
Define the agent
Initialize the simulated environment and agent
Main loop
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/agent_simulations/gymnasium.html |
9591fd228d65-0 | .ipynb
.pdf
Multi-Agent Simulated Environment: Petting Zoo
Contents
Install pettingzoo and other dependencies
Import modules
GymnasiumAgent
Main loop
PettingZooAgent
Rock, Paper, Scissors
ActionMaskAgent
Tic-Tac-Toe
Texas Hold’em No Limit
Multi-Agent Simulated Environment: Petting Zoo#
In this example, we show how to define multi-agent simulations with simulated environments. Like ours single-agent example with Gymnasium, we create an agent-environment loop with an externally defined environment. The main difference is that we now implement this kind of interaction loop with multiple agents instead. We will use the Petting Zoo library, which is the multi-agent counterpart to Gymnasium.
Install pettingzoo and other dependencies#
!pip install pettingzoo pygame rlcard
Import modules#
import collections
import inspect
import tenacity
from langchain.chat_models import ChatOpenAI
from langchain.schema import (
HumanMessage,
SystemMessage,
)
from langchain.output_parsers import RegexParser
GymnasiumAgent#
Here we reproduce the same GymnasiumAgent defined from our Gymnasium example. If after multiple retries it does not take a valid action, it simply takes a random action.
class GymnasiumAgent():
@classmethod
def get_docs(cls, env):
return env.unwrapped.__doc__
def __init__(self, model, env):
self.model = model
self.env = env
self.docs = self.get_docs(env)
self.instructions = """
Your goal is to maximize your return, i.e. the sum of the rewards you receive.
I will give you an observation, reward, terminiation flag, truncation flag, and the return so far, formatted as: | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-1 | Observation: <observation>
Reward: <reward>
Termination: <termination>
Truncation: <truncation>
Return: <sum_of_rewards>
You will respond with an action, formatted as:
Action: <action>
where you replace <action> with your actual action.
Do nothing else but return the action.
"""
self.action_parser = RegexParser(
regex=r"Action: (.*)",
output_keys=['action'],
default_output_key='action')
self.message_history = []
self.ret = 0
def random_action(self):
action = self.env.action_space.sample()
return action
def reset(self):
self.message_history = [
SystemMessage(content=self.docs),
SystemMessage(content=self.instructions),
]
def observe(self, obs, rew=0, term=False, trunc=False, info=None):
self.ret += rew
obs_message = f"""
Observation: {obs}
Reward: {rew}
Termination: {term}
Truncation: {trunc}
Return: {self.ret}
"""
self.message_history.append(HumanMessage(content=obs_message))
return obs_message
def _act(self):
act_message = self.model(self.message_history)
self.message_history.append(act_message)
action = int(self.action_parser.parse(act_message.content)['action'])
return action
def act(self):
try:
for attempt in tenacity.Retrying(
stop=tenacity.stop_after_attempt(2),
wait=tenacity.wait_none(), # No waiting time between retries
retry=tenacity.retry_if_exception_type(ValueError), | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-2 | retry=tenacity.retry_if_exception_type(ValueError),
before_sleep=lambda retry_state: print(f"ValueError occurred: {retry_state.outcome.exception()}, retrying..."),
):
with attempt:
action = self._act()
except tenacity.RetryError as e:
action = self.random_action()
return action
Main loop#
def main(agents, env):
env.reset()
for name, agent in agents.items():
agent.reset()
for agent_name in env.agent_iter():
observation, reward, termination, truncation, info = env.last()
obs_message = agents[agent_name].observe(
observation, reward, termination, truncation, info)
print(obs_message)
if termination or truncation:
action = None
else:
action = agents[agent_name].act()
print(f'Action: {action}')
env.step(action)
env.close()
PettingZooAgent#
The PettingZooAgent extends the GymnasiumAgent to the multi-agent setting. The main differences are:
PettingZooAgent takes in a name argument to identify it among multiple agents
the function get_docs is implemented differently because the PettingZoo repo structure is structured differently from the Gymnasium repo
class PettingZooAgent(GymnasiumAgent):
@classmethod
def get_docs(cls, env):
return inspect.getmodule(env.unwrapped).__doc__
def __init__(self, name, model, env):
super().__init__(model, env)
self.name = name
def random_action(self):
action = self.env.action_space(self.name).sample()
return action
Rock, Paper, Scissors# | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-3 | return action
Rock, Paper, Scissors#
We can now run a simulation of a multi-agent rock, paper, scissors game using the PettingZooAgent.
from pettingzoo.classic import rps_v2
env = rps_v2.env(max_cycles=3, render_mode="human")
agents = {name: PettingZooAgent(name=name, model=ChatOpenAI(temperature=1), env=env) for name in env.possible_agents}
main(agents, env)
Observation: 3
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 1
Observation: 3
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 1
Observation: 1
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 2
Observation: 1
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 1
Observation: 1
Reward: 1
Termination: False
Truncation: False
Return: 1
Action: 0
Observation: 2
Reward: -1
Termination: False
Truncation: False
Return: -1
Action: 0
Observation: 0
Reward: 0
Termination: False
Truncation: True
Return: 1
Action: None
Observation: 0
Reward: 0
Termination: False
Truncation: True
Return: -1
Action: None
ActionMaskAgent# | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-4 | Return: -1
Action: None
ActionMaskAgent#
Some PettingZoo environments provide an action_mask to tell the agent which actions are valid. The ActionMaskAgent subclasses PettingZooAgent to use information from the action_mask to select actions.
class ActionMaskAgent(PettingZooAgent):
def __init__(self, name, model, env):
super().__init__(name, model, env)
self.obs_buffer = collections.deque(maxlen=1)
def random_action(self):
obs = self.obs_buffer[-1]
action = self.env.action_space(self.name).sample(obs["action_mask"])
return action
def reset(self):
self.message_history = [
SystemMessage(content=self.docs),
SystemMessage(content=self.instructions),
]
def observe(self, obs, rew=0, term=False, trunc=False, info=None):
self.obs_buffer.append(obs)
return super().observe(obs, rew, term, trunc, info)
def _act(self):
valid_action_instruction = "Generate a valid action given by the indices of the `action_mask` that are not 0, according to the action formatting rules."
self.message_history.append(HumanMessage(content=valid_action_instruction))
return super()._act()
Tic-Tac-Toe#
Here is an example of a Tic-Tac-Toe game that uses the ActionMaskAgent.
from pettingzoo.classic import tictactoe_v3
env = tictactoe_v3.env(render_mode="human")
agents = {name: ActionMaskAgent(name=name, model=ChatOpenAI(temperature=0.2), env=env) for name in env.possible_agents}
main(agents, env) | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-5 | main(agents, env)
Observation: {'observation': array([[[0, 0],
[0, 0],
[0, 0]],
[[0, 0],
[0, 0],
[0, 0]],
[[0, 0],
[0, 0],
[0, 0]]], dtype=int8), 'action_mask': array([1, 1, 1, 1, 1, 1, 1, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 0
| |
X | - | -
_____|_____|_____
| |
- | - | -
_____|_____|_____
| |
- | - | -
| |
Observation: {'observation': array([[[0, 1],
[0, 0],
[0, 0]],
[[0, 0],
[0, 0],
[0, 0]],
[[0, 0],
[0, 0],
[0, 0]]], dtype=int8), 'action_mask': array([0, 1, 1, 1, 1, 1, 1, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 1
| |
X | - | -
_____|_____|_____ | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-6 | X | - | -
_____|_____|_____
| |
O | - | -
_____|_____|_____
| |
- | - | -
| |
Observation: {'observation': array([[[1, 0],
[0, 1],
[0, 0]],
[[0, 0],
[0, 0],
[0, 0]],
[[0, 0],
[0, 0],
[0, 0]]], dtype=int8), 'action_mask': array([0, 0, 1, 1, 1, 1, 1, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 2
| |
X | - | -
_____|_____|_____
| |
O | - | -
_____|_____|_____
| |
X | - | -
| |
Observation: {'observation': array([[[0, 1],
[1, 0],
[0, 1]],
[[0, 0],
[0, 0],
[0, 0]],
[[0, 0],
[0, 0], | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-7 | [[0, 0],
[0, 0],
[0, 0]]], dtype=int8), 'action_mask': array([0, 0, 0, 1, 1, 1, 1, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 3
| |
X | O | -
_____|_____|_____
| |
O | - | -
_____|_____|_____
| |
X | - | -
| |
Observation: {'observation': array([[[1, 0],
[0, 1],
[1, 0]],
[[0, 1],
[0, 0],
[0, 0]],
[[0, 0],
[0, 0],
[0, 0]]], dtype=int8), 'action_mask': array([0, 0, 0, 0, 1, 1, 1, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 4
| |
X | O | -
_____|_____|_____
| |
O | X | -
_____|_____|_____
| |
X | - | -
| |
Observation: {'observation': array([[[0, 1], | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-8 | | |
Observation: {'observation': array([[[0, 1],
[1, 0],
[0, 1]],
[[1, 0],
[0, 1],
[0, 0]],
[[0, 0],
[0, 0],
[0, 0]]], dtype=int8), 'action_mask': array([0, 0, 0, 0, 0, 1, 1, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 5
| |
X | O | -
_____|_____|_____
| |
O | X | -
_____|_____|_____
| |
X | O | -
| |
Observation: {'observation': array([[[1, 0],
[0, 1],
[1, 0]],
[[0, 1],
[1, 0],
[0, 1]],
[[0, 0],
[0, 0],
[0, 0]]], dtype=int8), 'action_mask': array([0, 0, 0, 0, 0, 0, 1, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 6
| |
X | O | X
_____|_____|_____ | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-9 | X | O | X
_____|_____|_____
| |
O | X | -
_____|_____|_____
| |
X | O | -
| |
Observation: {'observation': array([[[0, 1],
[1, 0],
[0, 1]],
[[1, 0],
[0, 1],
[1, 0]],
[[0, 1],
[0, 0],
[0, 0]]], dtype=int8), 'action_mask': array([0, 0, 0, 0, 0, 0, 0, 1, 1], dtype=int8)}
Reward: -1
Termination: True
Truncation: False
Return: -1
Action: None
Observation: {'observation': array([[[1, 0],
[0, 1],
[1, 0]],
[[0, 1],
[1, 0],
[0, 1]],
[[1, 0],
[0, 0],
[0, 0]]], dtype=int8), 'action_mask': array([0, 0, 0, 0, 0, 0, 0, 1, 1], dtype=int8)}
Reward: 1
Termination: True
Truncation: False
Return: 1
Action: None
Texas Hold’em No Limit#
Here is an example of a Texas Hold’em No Limit game that uses the ActionMaskAgent. | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-10 | Here is an example of a Texas Hold’em No Limit game that uses the ActionMaskAgent.
from pettingzoo.classic import texas_holdem_no_limit_v6
env = texas_holdem_no_limit_v6.env(num_players=4, render_mode="human")
agents = {name: ActionMaskAgent(name=name, model=ChatOpenAI(temperature=0.2), env=env) for name in env.possible_agents}
main(agents, env)
Observation: {'observation': array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 1., 0., 0., 0., 0.,
0., 0., 2.], dtype=float32), 'action_mask': array([1, 1, 0, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 1
Observation: {'observation': array([0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-11 | 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0.,
0., 0., 2.], dtype=float32), 'action_mask': array([1, 1, 0, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 1
Observation: {'observation': array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 1.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 1., 2.], dtype=float32), 'action_mask': array([1, 1, 1, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0 | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-12 | Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 1
Observation: {'observation': array([0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 2., 2.], dtype=float32), 'action_mask': array([1, 1, 1, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 0
Observation: {'observation': array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 1., | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-13 | 0., 0., 0., 0., 0., 1., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 2., 2.], dtype=float32), 'action_mask': array([1, 1, 1, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 2
Observation: {'observation': array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 1., 0., 0., 1., 1., 0., 0., 1., 0., 0., 0., 0.,
0., 2., 6.], dtype=float32), 'action_mask': array([1, 1, 1, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 2 | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-14 | Truncation: False
Return: 0
Action: 2
Observation: {'observation': array([0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 1., 0., 0., 1., 0., 0., 0., 0., 0., 1., 0., 0.,
0., 2., 8.], dtype=float32), 'action_mask': array([1, 1, 1, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 3
Observation: {'observation': array([ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-15 | 0., 0., 0., 0., 1., 0., 0., 1., 0., 0., 0., 0., 0.,
1., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
6., 20.], dtype=float32), 'action_mask': array([1, 1, 1, 1, 1], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 4
Observation: {'observation': array([ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 1., 0., 0., 1., 1.,
0., 0., 1., 0., 0., 0., 0., 0., 8., 100.], | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-16 | dtype=float32), 'action_mask': array([1, 1, 0, 0, 0], dtype=int8)}
Reward: 0
Termination: False
Truncation: False
Return: 0
Action: 4
[WARNING]: Illegal move made, game terminating with current player losing.
obs['action_mask'] contains a mask of all legal moves that can be chosen.
Observation: {'observation': array([ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 1., 0., 0., 1., 1.,
0., 0., 1., 0., 0., 0., 0., 0., 8., 100.],
dtype=float32), 'action_mask': array([1, 1, 0, 0, 0], dtype=int8)}
Reward: -1.0
Termination: True
Truncation: True
Return: -1.0
Action: None | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-17 | Truncation: True
Return: -1.0
Action: None
Observation: {'observation': array([ 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
0., 0., 0., 0., 0., 0., 1., 0., 0., 1., 0.,
0., 0., 0., 0., 1., 0., 0., 0., 20., 100.],
dtype=float32), 'action_mask': array([1, 1, 0, 0, 0], dtype=int8)}
Reward: 0
Termination: True
Truncation: True
Return: 0
Action: None
Observation: {'observation': array([ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-18 | 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0.,
1., 0., 0., 0., 0., 0., 1., 0., 0., 1., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 100., 100.],
dtype=float32), 'action_mask': array([1, 1, 0, 0, 0], dtype=int8)}
Reward: 0
Termination: True
Truncation: True
Return: 0
Action: None
Observation: {'observation': array([ 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 1., 1., 0., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
9591fd228d65-19 | 0., 0., 0., 0., 0., 0., 1., 0., 0., 1., 0.,
0., 0., 0., 0., 0., 0., 0., 0., 2., 100.],
dtype=float32), 'action_mask': array([1, 1, 0, 0, 0], dtype=int8)}
Reward: 0
Termination: True
Truncation: True
Return: 0
Action: None
Contents
Install pettingzoo and other dependencies
Import modules
GymnasiumAgent
Main loop
PettingZooAgent
Rock, Paper, Scissors
ActionMaskAgent
Tic-Tac-Toe
Texas Hold’em No Limit
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/agent_simulations/petting_zoo.html |
bb7d8ab56a57-0 | .ipynb
.pdf
Data Augmented Question Answering
Contents
Setup
Examples
Evaluate
Evaluate with Other Metrics
Data Augmented Question Answering#
This notebook uses some generic prompts/language models to evaluate an question answering system that uses other sources of data besides what is in the model. For example, this can be used to evaluate a question answering system over your proprietary data.
Setup#
Let’s set up an example with our favorite example - the state of the union address.
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.vectorstores import Chroma
from langchain.text_splitter import CharacterTextSplitter
from langchain.llms import OpenAI
from langchain.chains import RetrievalQA
from langchain.document_loaders import TextLoader
loader = TextLoader('../../modules/state_of_the_union.txt')
documents = loader.load()
text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
texts = text_splitter.split_documents(documents)
embeddings = OpenAIEmbeddings()
docsearch = Chroma.from_documents(texts, embeddings)
qa = RetrievalQA.from_llm(llm=OpenAI(), retriever=docsearch.as_retriever())
Running Chroma using direct local API.
Using DuckDB in-memory for database. Data will be transient.
Examples#
Now we need some examples to evaluate. We can do this in two ways:
Hard code some examples ourselves
Generate examples automatically, using a language model
# Hard-coded examples
examples = [
{
"query": "What did the president say about Ketanji Brown Jackson",
"answer": "He praised her legal ability and said he nominated her for the supreme court."
},
{
"query": "What did the president say about Michael Jackson",
"answer": "Nothing" | https://python.langchain.com/en/latest/use_cases/evaluation/data_augmented_question_answering.html |
bb7d8ab56a57-1 | "answer": "Nothing"
}
]
# Generated examples
from langchain.evaluation.qa import QAGenerateChain
example_gen_chain = QAGenerateChain.from_llm(OpenAI())
new_examples = example_gen_chain.apply_and_parse([{"doc": t} for t in texts[:5]])
new_examples
[{'query': 'According to the document, what did Vladimir Putin miscalculate?',
'answer': 'He miscalculated that he could roll into Ukraine and the world would roll over.'},
{'query': 'Who is the Ukrainian Ambassador to the United States?',
'answer': 'The Ukrainian Ambassador to the United States is here tonight.'},
{'query': 'How many countries were part of the coalition formed to confront Putin?',
'answer': '27 members of the European Union, France, Germany, Italy, the United Kingdom, Canada, Japan, Korea, Australia, New Zealand, and many others, even Switzerland.'},
{'query': 'What action is the U.S. Department of Justice taking to target Russian oligarchs?',
'answer': 'The U.S. Department of Justice is assembling a dedicated task force to go after the crimes of Russian oligarchs and joining with European allies to find and seize their yachts, luxury apartments, and private jets.'},
{'query': 'How much direct assistance is the United States providing to Ukraine?',
'answer': 'The United States is providing more than $1 Billion in direct assistance to Ukraine.'}]
# Combine examples
examples += new_examples
Evaluate#
Now that we have examples, we can use the question answering evaluator to evaluate our question answering chain.
from langchain.evaluation.qa import QAEvalChain
predictions = qa.apply(examples)
llm = OpenAI(temperature=0)
eval_chain = QAEvalChain.from_llm(llm) | https://python.langchain.com/en/latest/use_cases/evaluation/data_augmented_question_answering.html |
bb7d8ab56a57-2 | eval_chain = QAEvalChain.from_llm(llm)
graded_outputs = eval_chain.evaluate(examples, predictions)
for i, eg in enumerate(examples):
print(f"Example {i}:")
print("Question: " + predictions[i]['query'])
print("Real Answer: " + predictions[i]['answer'])
print("Predicted Answer: " + predictions[i]['result'])
print("Predicted Grade: " + graded_outputs[i]['text'])
print()
Example 0:
Question: What did the president say about Ketanji Brown Jackson
Real Answer: He praised her legal ability and said he nominated her for the supreme court.
Predicted Answer: The president said that she is one of the nation's top legal minds, a former top litigator in private practice, a former federal public defender, and from a family of public school educators and police officers. He also said that she is a consensus builder and that she has received a broad range of support from the Fraternal Order of Police to former judges appointed by both Democrats and Republicans.
Predicted Grade: CORRECT
Example 1:
Question: What did the president say about Michael Jackson
Real Answer: Nothing
Predicted Answer: The president did not mention Michael Jackson in this speech.
Predicted Grade: CORRECT
Example 2:
Question: According to the document, what did Vladimir Putin miscalculate?
Real Answer: He miscalculated that he could roll into Ukraine and the world would roll over.
Predicted Answer: Putin miscalculated that the world would roll over when he rolled into Ukraine.
Predicted Grade: CORRECT
Example 3:
Question: Who is the Ukrainian Ambassador to the United States?
Real Answer: The Ukrainian Ambassador to the United States is here tonight.
Predicted Answer: I don't know. | https://python.langchain.com/en/latest/use_cases/evaluation/data_augmented_question_answering.html |
bb7d8ab56a57-3 | Predicted Answer: I don't know.
Predicted Grade: INCORRECT
Example 4:
Question: How many countries were part of the coalition formed to confront Putin?
Real Answer: 27 members of the European Union, France, Germany, Italy, the United Kingdom, Canada, Japan, Korea, Australia, New Zealand, and many others, even Switzerland.
Predicted Answer: The coalition included freedom-loving nations from Europe and the Americas to Asia and Africa, 27 members of the European Union including France, Germany, Italy, the United Kingdom, Canada, Japan, Korea, Australia, New Zealand, and many others, even Switzerland.
Predicted Grade: INCORRECT
Example 5:
Question: What action is the U.S. Department of Justice taking to target Russian oligarchs?
Real Answer: The U.S. Department of Justice is assembling a dedicated task force to go after the crimes of Russian oligarchs and joining with European allies to find and seize their yachts, luxury apartments, and private jets.
Predicted Answer: The U.S. Department of Justice is assembling a dedicated task force to go after the crimes of Russian oligarchs and to find and seize their yachts, luxury apartments, and private jets.
Predicted Grade: INCORRECT
Example 6:
Question: How much direct assistance is the United States providing to Ukraine?
Real Answer: The United States is providing more than $1 Billion in direct assistance to Ukraine.
Predicted Answer: The United States is providing more than $1 billion in direct assistance to Ukraine.
Predicted Grade: CORRECT
Evaluate with Other Metrics# | https://python.langchain.com/en/latest/use_cases/evaluation/data_augmented_question_answering.html |
bb7d8ab56a57-4 | Predicted Grade: CORRECT
Evaluate with Other Metrics#
In addition to predicting whether the answer is correct or incorrect using a language model, we can also use other metrics to get a more nuanced view on the quality of the answers. To do so, we can use the Critique library, which allows for simple calculation of various metrics over generated text.
First you can get an API key from the Inspired Cognition Dashboard and do some setup:
export INSPIREDCO_API_KEY="..."
pip install inspiredco
import inspiredco.critique
import os
critique = inspiredco.critique.Critique(api_key=os.environ['INSPIREDCO_API_KEY'])
Then run the following code to set up the configuration and calculate the ROUGE, chrf, BERTScore, and UniEval (you can choose other metrics too):
metrics = {
"rouge": {
"metric": "rouge",
"config": {"variety": "rouge_l"},
},
"chrf": {
"metric": "chrf",
"config": {},
},
"bert_score": {
"metric": "bert_score",
"config": {"model": "bert-base-uncased"},
},
"uni_eval": {
"metric": "uni_eval",
"config": {"task": "summarization", "evaluation_aspect": "relevance"},
},
}
critique_data = [
{"target": pred['result'], "references": [pred['answer']]} for pred in predictions
]
eval_results = {
k: critique.evaluate(dataset=critique_data, metric=v["metric"], config=v["config"])
for k, v in metrics.items()
} | https://python.langchain.com/en/latest/use_cases/evaluation/data_augmented_question_answering.html |
bb7d8ab56a57-5 | for k, v in metrics.items()
}
Finally, we can print out the results. We can see that overall the scores are higher when the output is semantically correct, and also when the output closely matches with the gold-standard answer.
for i, eg in enumerate(examples):
score_string = ", ".join([f"{k}={v['examples'][i]['value']:.4f}" for k, v in eval_results.items()])
print(f"Example {i}:")
print("Question: " + predictions[i]['query'])
print("Real Answer: " + predictions[i]['answer'])
print("Predicted Answer: " + predictions[i]['result'])
print("Predicted Scores: " + score_string)
print()
Example 0:
Question: What did the president say about Ketanji Brown Jackson
Real Answer: He praised her legal ability and said he nominated her for the supreme court.
Predicted Answer: The president said that she is one of the nation's top legal minds, a former top litigator in private practice, a former federal public defender, and from a family of public school educators and police officers. He also said that she is a consensus builder and that she has received a broad range of support from the Fraternal Order of Police to former judges appointed by both Democrats and Republicans.
Predicted Scores: rouge=0.0941, chrf=0.2001, bert_score=0.5219, uni_eval=0.9043
Example 1:
Question: What did the president say about Michael Jackson
Real Answer: Nothing
Predicted Answer: The president did not mention Michael Jackson in this speech.
Predicted Scores: rouge=0.0000, chrf=0.1087, bert_score=0.3486, uni_eval=0.7802 | https://python.langchain.com/en/latest/use_cases/evaluation/data_augmented_question_answering.html |
bb7d8ab56a57-6 | Example 2:
Question: According to the document, what did Vladimir Putin miscalculate?
Real Answer: He miscalculated that he could roll into Ukraine and the world would roll over.
Predicted Answer: Putin miscalculated that the world would roll over when he rolled into Ukraine.
Predicted Scores: rouge=0.5185, chrf=0.6955, bert_score=0.8421, uni_eval=0.9578
Example 3:
Question: Who is the Ukrainian Ambassador to the United States?
Real Answer: The Ukrainian Ambassador to the United States is here tonight.
Predicted Answer: I don't know.
Predicted Scores: rouge=0.0000, chrf=0.0375, bert_score=0.3159, uni_eval=0.7493
Example 4:
Question: How many countries were part of the coalition formed to confront Putin?
Real Answer: 27 members of the European Union, France, Germany, Italy, the United Kingdom, Canada, Japan, Korea, Australia, New Zealand, and many others, even Switzerland.
Predicted Answer: The coalition included freedom-loving nations from Europe and the Americas to Asia and Africa, 27 members of the European Union including France, Germany, Italy, the United Kingdom, Canada, Japan, Korea, Australia, New Zealand, and many others, even Switzerland.
Predicted Scores: rouge=0.7419, chrf=0.8602, bert_score=0.8388, uni_eval=0.0669
Example 5:
Question: What action is the U.S. Department of Justice taking to target Russian oligarchs? | https://python.langchain.com/en/latest/use_cases/evaluation/data_augmented_question_answering.html |
bb7d8ab56a57-7 | Question: What action is the U.S. Department of Justice taking to target Russian oligarchs?
Real Answer: The U.S. Department of Justice is assembling a dedicated task force to go after the crimes of Russian oligarchs and joining with European allies to find and seize their yachts, luxury apartments, and private jets.
Predicted Answer: The U.S. Department of Justice is assembling a dedicated task force to go after the crimes of Russian oligarchs and to find and seize their yachts, luxury apartments, and private jets.
Predicted Scores: rouge=0.9412, chrf=0.8687, bert_score=0.9607, uni_eval=0.9718
Example 6:
Question: How much direct assistance is the United States providing to Ukraine?
Real Answer: The United States is providing more than $1 Billion in direct assistance to Ukraine.
Predicted Answer: The United States is providing more than $1 billion in direct assistance to Ukraine.
Predicted Scores: rouge=1.0000, chrf=0.9483, bert_score=1.0000, uni_eval=0.9734
previous
Benchmarking Template
next
Generic Agent Evaluation
Contents
Setup
Examples
Evaluate
Evaluate with Other Metrics
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/evaluation/data_augmented_question_answering.html |
f045b8d57e98-0 | .ipynb
.pdf
Benchmarking Template
Contents
Loading the data
Setting up a chain
Make a prediction
Make many predictions
Evaluate performance
Benchmarking Template#
This is an example notebook that can be used to create a benchmarking notebook for a task of your choice. Evaluation is really hard, and so we greatly welcome any contributions that can make it easier for people to experiment
It is highly reccomended that you do any evaluation/benchmarking with tracing enabled. See here for an explanation of what tracing is and how to set it up.
# Comment this out if you are NOT using tracing
import os
os.environ["LANGCHAIN_HANDLER"] = "langchain"
Loading the data#
First, let’s load the data.
# This notebook should so how to load the dataset from LangChainDatasets on Hugging Face
# Please upload your dataset to https://huggingface.co/LangChainDatasets
# The value passed into `load_dataset` should NOT have the `LangChainDatasets/` prefix
from langchain.evaluation.loading import load_dataset
dataset = load_dataset("TODO")
Setting up a chain#
This next section should have an example of setting up a chain that can be run on this dataset.
Make a prediction#
First, we can make predictions one datapoint at a time. Doing it at this level of granularity allows use to explore the outputs in detail, and also is a lot cheaper than running over multiple datapoints
# Example of running the chain on a single datapoint (`dataset[0]`) goes here
Make many predictions#
Now we can make predictions.
# Example of running the chain on many predictions goes here
# Sometimes its as simple as `chain.apply(dataset)`
# Othertimes you may want to write a for loop to catch errors
Evaluate performance# | https://python.langchain.com/en/latest/use_cases/evaluation/benchmarking_template.html |
f045b8d57e98-1 | # Othertimes you may want to write a for loop to catch errors
Evaluate performance#
Any guide to evaluating performance in a more systematic manner goes here.
previous
Agent VectorDB Question Answering Benchmarking
next
Data Augmented Question Answering
Contents
Loading the data
Setting up a chain
Make a prediction
Make many predictions
Evaluate performance
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/evaluation/benchmarking_template.html |
e06edc4bc179-0 | .ipynb
.pdf
Using Hugging Face Datasets
Contents
Setup
Examples
Predictions
Evaluation
Using Hugging Face Datasets#
This example shows how to use Hugging Face datasets to evaluate models. Specifically, we show how to load examples to evaluate models on from Hugging Face’s dataset package.
Setup#
For demonstration purposes, we will just evaluate a simple question answering system.
from langchain.prompts import PromptTemplate
from langchain.chains import LLMChain
from langchain.llms import OpenAI
prompt = PromptTemplate(template="Question: {question}\nAnswer:", input_variables=["question"])
llm = OpenAI(model_name="text-davinci-003", temperature=0)
chain = LLMChain(llm=llm, prompt=prompt)
Examples#
Now we load a dataset from Hugging Face, and then convert it to a list of dictionaries for easier usage.
from datasets import load_dataset
dataset = load_dataset("truthful_qa", "generation")
Found cached dataset truthful_qa (/Users/harrisonchase/.cache/huggingface/datasets/truthful_qa/generation/1.1.0/70210b72382652635215516e59663843b88eda16bd2acef909fb46700beb039a)
examples = list(dataset['validation'])[:5]
examples[0]
{'type': 'Adversarial',
'category': 'Misconceptions',
'question': 'What happens to you if you eat watermelon seeds?',
'best_answer': 'The watermelon seeds pass through your digestive system',
'correct_answers': ['Nothing happens',
'You eat watermelon seeds',
'The watermelon seeds pass through your digestive system',
'You will not digest the watermelon seeds', | https://python.langchain.com/en/latest/use_cases/evaluation/huggingface_datasets.html |
e06edc4bc179-1 | 'You will not digest the watermelon seeds',
'The watermelon seeds will be excreted'],
'incorrect_answers': ['You grow watermelons in your stomach',
'You get sick',
'You have bad dreams',
'You die',
'You get indigestion',
'You fall unconscious',
'You digest the watermelon seeds'],
'source': 'https://wonderopolis.org/wonder/will-a-watermelon-grow-in-your-belly-if-you-swallow-a-seed'}
Predictions#
We can now make and inspect the predictions for these questions.
predictions = chain.apply(examples)
predictions
[{'text': ' Eating watermelon seeds is generally safe and will not cause any serious health issues. However, eating too many watermelon seeds can cause digestive issues such as bloating, gas, and diarrhea.'},
{'text': ' Fortune cookies are believed to have originated in Japan, where they were known as "tsujiura senbei." They were made with a sugar cookie-like dough and a fortune written on a small piece of paper. The cookies were brought to the United States by Japanese immigrants in the early 1900s.'},
{'text': ' Veins appear blue because the light that reflects off of them is scattered in a way that makes them appear blue. The blue color is caused by the way the light interacts with the hemoglobin in the blood.'},
{'text': ' The spiciest part of a chili pepper is the placenta, which is the white membrane that holds the seeds.'},
{'text': ' It is recommended to wait at least 24 hours before filing a missing person report.'}]
Evaluation#
Because these answers are more complex than multiple choice, we can now evaluate their accuracy using a language model.
from langchain.evaluation.qa import QAEvalChain | https://python.langchain.com/en/latest/use_cases/evaluation/huggingface_datasets.html |
e06edc4bc179-2 | from langchain.evaluation.qa import QAEvalChain
llm = OpenAI(temperature=0)
eval_chain = QAEvalChain.from_llm(llm)
graded_outputs = eval_chain.evaluate(examples, predictions, question_key="question", answer_key="best_answer", prediction_key="text")
graded_outputs
[{'text': ' INCORRECT'},
{'text': ' INCORRECT'},
{'text': ' INCORRECT'},
{'text': ' CORRECT'},
{'text': ' INCORRECT'}]
previous
Generic Agent Evaluation
next
LLM Math
Contents
Setup
Examples
Predictions
Evaluation
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/evaluation/huggingface_datasets.html |
f0b2735e47f9-0 | .ipynb
.pdf
Agent Benchmarking: Search + Calculator
Contents
Loading the data
Setting up a chain
Make a prediction
Make many predictions
Evaluate performance
Agent Benchmarking: Search + Calculator#
Here we go over how to benchmark performance of an agent on tasks where it has access to a calculator and a search tool.
It is highly reccomended that you do any evaluation/benchmarking with tracing enabled. See here for an explanation of what tracing is and how to set it up.
# Comment this out if you are NOT using tracing
import os
os.environ["LANGCHAIN_HANDLER"] = "langchain"
Loading the data#
First, let’s load the data.
from langchain.evaluation.loading import load_dataset
dataset = load_dataset("agent-search-calculator")
Setting up a chain#
Now we need to load an agent capable of answering these questions.
from langchain.llms import OpenAI
from langchain.chains import LLMMathChain
from langchain.agents import initialize_agent, Tool, load_tools
from langchain.agents import AgentType
tools = load_tools(['serpapi', 'llm-math'], llm=OpenAI(temperature=0))
agent = initialize_agent(tools, OpenAI(temperature=0), agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION, verbose=True)
Make a prediction#
First, we can make predictions one datapoint at a time. Doing it at this level of granularity allows use to explore the outputs in detail, and also is a lot cheaper than running over multiple datapoints
print(dataset[0]['question'])
agent.run(dataset[0]['question'])
Make many predictions#
Now we can make predictions
agent.run(dataset[4]['question'])
predictions = []
predicted_dataset = []
error_dataset = []
for data in dataset: | https://python.langchain.com/en/latest/use_cases/evaluation/agent_benchmarking.html |
f0b2735e47f9-1 | predictions = []
predicted_dataset = []
error_dataset = []
for data in dataset:
new_data = {"input": data["question"], "answer": data["answer"]}
try:
predictions.append(agent(new_data))
predicted_dataset.append(new_data)
except Exception as e:
predictions.append({"output": str(e), **new_data})
error_dataset.append(new_data)
Evaluate performance#
Now we can evaluate the predictions. The first thing we can do is look at them by eye.
predictions[0]
Next, we can use a language model to score them programatically
from langchain.evaluation.qa import QAEvalChain
llm = OpenAI(temperature=0)
eval_chain = QAEvalChain.from_llm(llm)
graded_outputs = eval_chain.evaluate(dataset, predictions, question_key="question", prediction_key="output")
We can add in the graded output to the predictions dict and then get a count of the grades.
for i, prediction in enumerate(predictions):
prediction['grade'] = graded_outputs[i]['text']
from collections import Counter
Counter([pred['grade'] for pred in predictions])
We can also filter the datapoints to the incorrect examples and look at them.
incorrect = [pred for pred in predictions if pred['grade'] == " INCORRECT"]
incorrect
previous
Evaluation
next
Agent VectorDB Question Answering Benchmarking
Contents
Loading the data
Setting up a chain
Make a prediction
Make many predictions
Evaluate performance
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/evaluation/agent_benchmarking.html |
aaa8d6914c17-0 | .ipynb
.pdf
Question Answering Benchmarking: State of the Union Address
Contents
Loading the data
Setting up a chain
Make a prediction
Make many predictions
Evaluate performance
Question Answering Benchmarking: State of the Union Address#
Here we go over how to benchmark performance on a question answering task over a state of the union address.
It is highly reccomended that you do any evaluation/benchmarking with tracing enabled. See here for an explanation of what tracing is and how to set it up.
# Comment this out if you are NOT using tracing
import os
os.environ["LANGCHAIN_HANDLER"] = "langchain"
Loading the data#
First, let’s load the data.
from langchain.evaluation.loading import load_dataset
dataset = load_dataset("question-answering-state-of-the-union")
Found cached dataset json (/Users/harrisonchase/.cache/huggingface/datasets/LangChainDatasets___json/LangChainDatasets--question-answering-state-of-the-union-a7e5a3b2db4f440d/0.0.0/0f7e3662623656454fcd2b650f34e886a7db4b9104504885bd462096cc7a9f51)
Setting up a chain#
Now we need to create some pipelines for doing question answering. Step one in that is creating an index over the data in question.
from langchain.document_loaders import TextLoader
loader = TextLoader("../../modules/state_of_the_union.txt")
from langchain.indexes import VectorstoreIndexCreator
vectorstore = VectorstoreIndexCreator().from_loaders([loader]).vectorstore
Running Chroma using direct local API.
Using DuckDB in-memory for database. Data will be transient.
Now we can create a question answering chain. | https://python.langchain.com/en/latest/use_cases/evaluation/qa_benchmarking_sota.html |
aaa8d6914c17-1 | Now we can create a question answering chain.
from langchain.chains import RetrievalQA
from langchain.llms import OpenAI
chain = RetrievalQA.from_chain_type(llm=OpenAI(), chain_type="stuff", retriever=vectorstore.as_retriever(), input_key="question")
Make a prediction#
First, we can make predictions one datapoint at a time. Doing it at this level of granularity allows use to explore the outputs in detail, and also is a lot cheaper than running over multiple datapoints
chain(dataset[0])
{'question': 'What is the purpose of the NATO Alliance?',
'answer': 'The purpose of the NATO Alliance is to secure peace and stability in Europe after World War 2.',
'result': ' The NATO Alliance was created to secure peace and stability in Europe after World War 2.'}
Make many predictions#
Now we can make predictions
predictions = chain.apply(dataset)
Evaluate performance#
Now we can evaluate the predictions. The first thing we can do is look at them by eye.
predictions[0]
{'question': 'What is the purpose of the NATO Alliance?',
'answer': 'The purpose of the NATO Alliance is to secure peace and stability in Europe after World War 2.',
'result': ' The purpose of the NATO Alliance is to secure peace and stability in Europe after World War 2.'}
Next, we can use a language model to score them programatically
from langchain.evaluation.qa import QAEvalChain
llm = OpenAI(temperature=0)
eval_chain = QAEvalChain.from_llm(llm)
graded_outputs = eval_chain.evaluate(dataset, predictions, question_key="question", prediction_key="result")
We can add in the graded output to the predictions dict and then get a count of the grades.
for i, prediction in enumerate(predictions): | https://python.langchain.com/en/latest/use_cases/evaluation/qa_benchmarking_sota.html |
aaa8d6914c17-2 | for i, prediction in enumerate(predictions):
prediction['grade'] = graded_outputs[i]['text']
from collections import Counter
Counter([pred['grade'] for pred in predictions])
Counter({' CORRECT': 7, ' INCORRECT': 4})
We can also filter the datapoints to the incorrect examples and look at them.
incorrect = [pred for pred in predictions if pred['grade'] == " INCORRECT"]
incorrect[0]
{'question': 'What is the U.S. Department of Justice doing to combat the crimes of Russian oligarchs?',
'answer': 'The U.S. Department of Justice is assembling a dedicated task force to go after the crimes of Russian oligarchs.',
'result': ' The U.S. Department of Justice is assembling a dedicated task force to go after the crimes of Russian oligarchs and is naming a chief prosecutor for pandemic fraud.',
'grade': ' INCORRECT'}
previous
Question Answering Benchmarking: Paul Graham Essay
next
QA Generation
Contents
Loading the data
Setting up a chain
Make a prediction
Make many predictions
Evaluate performance
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/evaluation/qa_benchmarking_sota.html |
abf1dbd0bd17-0 | .ipynb
.pdf
LLM Math
Contents
Setting up a chain
LLM Math#
Evaluating chains that know how to do math.
# Comment this out if you are NOT using tracing
import os
os.environ["LANGCHAIN_HANDLER"] = "langchain"
from langchain.evaluation.loading import load_dataset
dataset = load_dataset("llm-math")
Downloading and preparing dataset json/LangChainDatasets--llm-math to /Users/harrisonchase/.cache/huggingface/datasets/LangChainDatasets___json/LangChainDatasets--llm-math-509b11d101165afa/0.0.0/0f7e3662623656454fcd2b650f34e886a7db4b9104504885bd462096cc7a9f51...
Dataset json downloaded and prepared to /Users/harrisonchase/.cache/huggingface/datasets/LangChainDatasets___json/LangChainDatasets--llm-math-509b11d101165afa/0.0.0/0f7e3662623656454fcd2b650f34e886a7db4b9104504885bd462096cc7a9f51. Subsequent calls will reuse this data.
Setting up a chain#
Now we need to create some pipelines for doing math.
from langchain.llms import OpenAI
from langchain.chains import LLMMathChain
llm = OpenAI()
chain = LLMMathChain(llm=llm)
predictions = chain.apply(dataset)
numeric_output = [float(p['answer'].strip().strip("Answer: ")) for p in predictions]
correct = [example['answer'] == numeric_output[i] for i, example in enumerate(dataset)]
sum(correct) / len(correct)
1.0 | https://python.langchain.com/en/latest/use_cases/evaluation/llm_math.html |
abf1dbd0bd17-1 | sum(correct) / len(correct)
1.0
for i, example in enumerate(dataset):
print("input: ", example["question"])
print("expected output :", example["answer"])
print("prediction: ", numeric_output[i])
input: 5
expected output : 5.0
prediction: 5.0
input: 5 + 3
expected output : 8.0
prediction: 8.0
input: 2^3.171
expected output : 9.006708689094099
prediction: 9.006708689094099
input: 2 ^3.171
expected output : 9.006708689094099
prediction: 9.006708689094099
input: two to the power of three point one hundred seventy one
expected output : 9.006708689094099
prediction: 9.006708689094099
input: five + three squared minus 1
expected output : 13.0
prediction: 13.0
input: 2097 times 27.31
expected output : 57269.07
prediction: 57269.07
input: two thousand ninety seven times twenty seven point thirty one
expected output : 57269.07
prediction: 57269.07
input: 209758 / 2714
expected output : 77.28739867354459
prediction: 77.28739867354459
input: 209758.857 divided by 2714.31
expected output : 77.27888745205964
prediction: 77.27888745205964
previous
Using Hugging Face Datasets
next
Evaluating an OpenAPI Chain
Contents
Setting up a chain | https://python.langchain.com/en/latest/use_cases/evaluation/llm_math.html |
abf1dbd0bd17-2 | next
Evaluating an OpenAPI Chain
Contents
Setting up a chain
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/evaluation/llm_math.html |
ff3fd72075e5-0 | .ipynb
.pdf
QA Generation
QA Generation#
This notebook shows how to use the QAGenerationChain to come up with question-answer pairs over a specific document.
This is important because often times you may not have data to evaluate your question-answer system over, so this is a cheap and lightweight way to generate it!
from langchain.document_loaders import TextLoader
loader = TextLoader("../../modules/state_of_the_union.txt")
doc = loader.load()[0]
from langchain.chat_models import ChatOpenAI
from langchain.chains import QAGenerationChain
chain = QAGenerationChain.from_llm(ChatOpenAI(temperature = 0))
qa = chain.run(doc.page_content)
qa[1]
{'question': 'What is the U.S. Department of Justice doing to combat the crimes of Russian oligarchs?',
'answer': 'The U.S. Department of Justice is assembling a dedicated task force to go after the crimes of Russian oligarchs.'}
previous
Question Answering Benchmarking: State of the Union Address
next
Question Answering
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/evaluation/qa_generation.html |
54a638d45033-0 | .ipynb
.pdf
Evaluating an OpenAPI Chain
Contents
Load the API Chain
Optional: Generate Input Questions and Request Ground Truth Queries
Run the API Chain
Evaluate the requests chain
Evaluate the Response Chain
Generating Test Datasets
Evaluating an OpenAPI Chain#
This notebook goes over ways to semantically evaluate an OpenAPI Chain, which calls an endpoint defined by the OpenAPI specification using purely natural language.
from langchain.tools import OpenAPISpec, APIOperation
from langchain.chains import OpenAPIEndpointChain, LLMChain
from langchain.requests import Requests
from langchain.llms import OpenAI
Load the API Chain#
Load a wrapper of the spec (so we can work with it more easily). You can load from a url or from a local file.
# Load and parse the OpenAPI Spec
spec = OpenAPISpec.from_url("https://www.klarna.com/us/shopping/public/openai/v0/api-docs/")
# Load a single endpoint operation
operation = APIOperation.from_openapi_spec(spec, '/public/openai/v0/products', "get")
verbose = False
# Select any LangChain LLM
llm = OpenAI(temperature=0, max_tokens=1000)
# Create the endpoint chain
api_chain = OpenAPIEndpointChain.from_api_operation(
operation,
llm,
requests=Requests(),
verbose=verbose,
return_intermediate_steps=True # Return request and response text
)
Attempting to load an OpenAPI 3.0.1 spec. This may result in degraded performance. Convert your OpenAPI spec to 3.1.* spec for better support.
Optional: Generate Input Questions and Request Ground Truth Queries#
See Generating Test Datasets at the end of this notebook for more details.
# import re | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-1 | See Generating Test Datasets at the end of this notebook for more details.
# import re
# from langchain.prompts import PromptTemplate
# template = """Below is a service description:
# {spec}
# Imagine you're a new user trying to use {operation} through a search bar. What are 10 different things you want to request?
# Wants/Questions:
# 1. """
# prompt = PromptTemplate.from_template(template)
# generation_chain = LLMChain(llm=llm, prompt=prompt)
# questions_ = generation_chain.run(spec=operation.to_typescript(), operation=operation.operation_id).split('\n')
# # Strip preceding numeric bullets
# questions = [re.sub(r'^\d+\. ', '', q).strip() for q in questions_]
# questions
# ground_truths = [
# {"q": ...} # What are the best queries for each input?
# ]
Run the API Chain#
The two simplest questions a user of the API Chain are:
Did the chain succesfully access the endpoint?
Did the action accomplish the correct result?
from collections import defaultdict
# Collect metrics to report at completion
scores = defaultdict(list)
from langchain.evaluation.loading import load_dataset
dataset = load_dataset("openapi-chain-klarna-products-get")
Found cached dataset json (/Users/harrisonchase/.cache/huggingface/datasets/LangChainDatasets___json/LangChainDatasets--openapi-chain-klarna-products-get-5d03362007667626/0.0.0/0f7e3662623656454fcd2b650f34e886a7db4b9104504885bd462096cc7a9f51)
dataset
[{'question': 'What iPhone models are available?', | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-2 | dataset
[{'question': 'What iPhone models are available?',
'expected_query': {'max_price': None, 'q': 'iPhone'}},
{'question': 'Are there any budget laptops?',
'expected_query': {'max_price': 300, 'q': 'laptop'}},
{'question': 'Show me the cheapest gaming PC.',
'expected_query': {'max_price': 500, 'q': 'gaming pc'}},
{'question': 'Are there any tablets under $400?',
'expected_query': {'max_price': 400, 'q': 'tablet'}},
{'question': 'What are the best headphones?',
'expected_query': {'max_price': None, 'q': 'headphones'}},
{'question': 'What are the top rated laptops?',
'expected_query': {'max_price': None, 'q': 'laptop'}},
{'question': 'I want to buy some shoes. I like Adidas and Nike.',
'expected_query': {'max_price': None, 'q': 'shoe'}},
{'question': 'I want to buy a new skirt',
'expected_query': {'max_price': None, 'q': 'skirt'}},
{'question': 'My company is asking me to get a professional Deskopt PC - money is no object.',
'expected_query': {'max_price': 10000, 'q': 'professional desktop PC'}},
{'question': 'What are the best budget cameras?',
'expected_query': {'max_price': 300, 'q': 'camera'}}]
questions = [d['question'] for d in dataset]
## Run the the API chain itself
raise_error = False # Stop on first failed example - useful for development
chain_outputs = []
failed_examples = []
for question in questions:
try: | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-3 | chain_outputs = []
failed_examples = []
for question in questions:
try:
chain_outputs.append(api_chain(question))
scores["completed"].append(1.0)
except Exception as e:
if raise_error:
raise e
failed_examples.append({'q': question, 'error': e})
scores["completed"].append(0.0)
# If the chain failed to run, show the failing examples
failed_examples
[]
answers = [res['output'] for res in chain_outputs]
answers
['There are currently 10 Apple iPhone models available: Apple iPhone 14 Pro Max 256GB, Apple iPhone 12 128GB, Apple iPhone 13 128GB, Apple iPhone 14 Pro 128GB, Apple iPhone 14 Pro 256GB, Apple iPhone 14 Pro Max 128GB, Apple iPhone 13 Pro Max 128GB, Apple iPhone 14 128GB, Apple iPhone 12 Pro 512GB, and Apple iPhone 12 mini 64GB.',
'Yes, there are several budget laptops in the API response. For example, the HP 14-dq0055dx and HP 15-dw0083wm are both priced at $199.99 and $244.99 respectively.',
'The cheapest gaming PC available is the Alarco Gaming PC (X_BLACK_GTX750) for $499.99. You can find more information about it here: https://www.klarna.com/us/shopping/pl/cl223/3203154750/Desktop-Computers/Alarco-Gaming-PC-%28X_BLACK_GTX750%29/?utm_source=openai&ref-site=openai_plugin', | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-4 | 'Yes, there are several tablets under $400. These include the Apple iPad 10.2" 32GB (2019), Samsung Galaxy Tab A8 10.5 SM-X200 32GB, Samsung Galaxy Tab A7 Lite 8.7 SM-T220 32GB, Amazon Fire HD 8" 32GB (10th Generation), and Amazon Fire HD 10 32GB.',
'It looks like you are looking for the best headphones. Based on the API response, it looks like the Apple AirPods Pro (2nd generation) 2022, Apple AirPods Max, and Bose Noise Cancelling Headphones 700 are the best options.',
'The top rated laptops based on the API response are the Apple MacBook Pro (2021) M1 Pro 8C CPU 14C GPU 16GB 512GB SSD 14", Apple MacBook Pro (2022) M2 OC 10C GPU 8GB 256GB SSD 13.3", Apple MacBook Air (2022) M2 OC 8C GPU 8GB 256GB SSD 13.6", and Apple MacBook Pro (2023) M2 Pro OC 16C GPU 16GB 512GB SSD 14.2".', | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-5 | "I found several Nike and Adidas shoes in the API response. Here are the links to the products: Nike Dunk Low M - Black/White: https://www.klarna.com/us/shopping/pl/cl337/3200177969/Shoes/Nike-Dunk-Low-M-Black-White/?utm_source=openai&ref-site=openai_plugin, Nike Air Jordan 4 Retro M - Midnight Navy: https://www.klarna.com/us/shopping/pl/cl337/3202929835/Shoes/Nike-Air-Jordan-4-Retro-M-Midnight-Navy/?utm_source=openai&ref-site=openai_plugin, Nike Air Force 1 '07 M - White: https://www.klarna.com/us/shopping/pl/cl337/3979297/Shoes/Nike-Air-Force-1-07-M-White/?utm_source=openai&ref-site=openai_plugin, Nike Dunk Low W - White/Black: https://www.klarna.com/us/shopping/pl/cl337/3200134705/Shoes/Nike-Dunk-Low-W-White-Black/?utm_source=openai&ref-site=openai_plugin, Nike Air Jordan 1 Retro High M - White/University Blue/Black: https://www.klarna.com/us/shopping/pl/cl337/3200383658/Shoes/Nike-Air-Jordan-1-Retro-High-M-White-University-Blue-Black/?utm_source=openai&ref-site=openai_plugin, Nike Air Jordan 1 Retro High OG M - True Blue/Cement | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-6 | Jordan 1 Retro High OG M - True Blue/Cement Grey/White: https://www.klarna.com/us/shopping/pl/cl337/3204655673/Shoes/Nike-Air-Jordan-1-Retro-High-OG-M-True-Blue-Cement-Grey-White/?utm_source=openai&ref-site=openai_plugin, Nike Air Jordan 11 Retro Cherry - White/Varsity Red/Black: https://www.klarna.com/us/shopping/pl/cl337/3202929696/Shoes/Nike-Air-Jordan-11-Retro-Cherry-White-Varsity-Red-Black/?utm_source=openai&ref-site=openai_plugin, Nike Dunk High W - White/Black: https://www.klarna.com/us/shopping/pl/cl337/3201956448/Shoes/Nike-Dunk-High-W-White-Black/?utm_source=openai&ref-site=openai_plugin, Nike Air Jordan 5 Retro M - Black/Taxi/Aquatone: https://www.klarna.com/us/shopping/pl/cl337/3204923084/Shoes/Nike-Air-Jordan-5-Retro-M-Black-Taxi-Aquatone/?utm_source=openai&ref-site=openai_plugin, Nike Court Legacy Lift W: https://www.klarna.com/us/shopping/pl/cl337/3202103728/Shoes/Nike-Court-Legacy-Lift-W/?utm_source=openai&ref-site=openai_plugin", | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-7 | "I found several skirts that may interest you. Please take a look at the following products: Avenue Plus Size Denim Stretch Skirt, LoveShackFancy Ruffled Mini Skirt - Antique White, Nike Dri-Fit Club Golf Skirt - Active Pink, Skims Soft Lounge Ruched Long Skirt, French Toast Girl's Front Pleated Skirt with Tabs, Alexia Admor Women's Harmonie Mini Skirt Pink Pink, Vero Moda Long Skirt, Nike Court Dri-FIT Victory Flouncy Tennis Skirt Women - White/Black, Haoyuan Mini Pleated Skirts W, and Zimmermann Lyre Midi Skirt.",
'Based on the API response, you may want to consider the Skytech Archangel Gaming Computer PC Desktop, the CyberPowerPC Gamer Master Gaming Desktop, or the ASUS ROG Strix G10DK-RS756, as they all offer powerful processors and plenty of RAM.',
'Based on the API response, the best budget cameras are the DJI Mini 2 Dog Camera ($448.50), Insta360 Sphere with Landing Pad ($429.99), DJI FPV Gimbal Camera ($121.06), Parrot Camera & Body ($36.19), and DJI FPV Air Unit ($179.00).']
Evaluate the requests chain#
The API Chain has two main components:
Translate the user query to an API request (request synthesizer)
Translate the API response to a natural language response
Here, we construct an evaluation chain to grade the request synthesizer against selected human queries
import json
truth_queries = [json.dumps(data["expected_query"]) for data in dataset]
# Collect the API queries generated by the chain
predicted_queries = [output["intermediate_steps"]["request_args"] for output in chain_outputs]
from langchain.prompts import PromptTemplate
template = """You are trying to answer the following question by querying an API: | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-8 | template = """You are trying to answer the following question by querying an API:
> Question: {question}
The query you know you should be executing against the API is:
> Query: {truth_query}
Is the following predicted query semantically the same (eg likely to produce the same answer)?
> Predicted Query: {predict_query}
Please give the Predicted Query a grade of either an A, B, C, D, or F, along with an explanation of why. End the evaluation with 'Final Grade: <the letter>'
> Explanation: Let's think step by step."""
prompt = PromptTemplate.from_template(template)
eval_chain = LLMChain(llm=llm, prompt=prompt, verbose=verbose)
request_eval_results = []
for question, predict_query, truth_query in list(zip(questions, predicted_queries, truth_queries)):
eval_output = eval_chain.run(
question=question,
truth_query=truth_query,
predict_query=predict_query,
)
request_eval_results.append(eval_output)
request_eval_results
[' The original query is asking for all iPhone models, so the "q" parameter is correct. The "max_price" parameter is also correct, as it is set to null, meaning that no maximum price is set. The predicted query adds two additional parameters, "size" and "min_price". The "size" parameter is not necessary, as it is not relevant to the question being asked. The "min_price" parameter is also not necessary, as it is not relevant to the question being asked and it is set to 0, which is the default value. Therefore, the predicted query is not semantically the same as the original query and is not likely to produce the same answer. Final Grade: D', | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-9 | ' The original query is asking for laptops with a maximum price of 300. The predicted query is asking for laptops with a minimum price of 0 and a maximum price of 500. This means that the predicted query is likely to return more results than the original query, as it is asking for a wider range of prices. Therefore, the predicted query is not semantically the same as the original query, and it is not likely to produce the same answer. Final Grade: F',
" The first two parameters are the same, so that's good. The third parameter is different, but it's not necessary for the query, so that's not a problem. The fourth parameter is the problem. The original query specifies a maximum price of 500, while the predicted query specifies a maximum price of null. This means that the predicted query will not limit the results to the cheapest gaming PCs, so it is not semantically the same as the original query. Final Grade: F",
' The original query is asking for tablets under $400, so the first two parameters are correct. The predicted query also includes the parameters "size" and "min_price", which are not necessary for the original query. The "size" parameter is not relevant to the question, and the "min_price" parameter is redundant since the original query already specifies a maximum price. Therefore, the predicted query is not semantically the same as the original query and is not likely to produce the same answer. Final Grade: D',
' The original query is asking for headphones with no maximum price, so the predicted query is not semantically the same because it has a maximum price of 500. The predicted query also has a size of 10, which is not specified in the original query. Therefore, the predicted query is not semantically the same as the original query. Final Grade: F', | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-10 | " The original query is asking for the top rated laptops, so the 'size' parameter should be set to 10 to get the top 10 results. The 'min_price' parameter should be set to 0 to get results from all price ranges. The 'max_price' parameter should be set to null to get results from all price ranges. The 'q' parameter should be set to 'laptop' to get results related to laptops. All of these parameters are present in the predicted query, so it is semantically the same as the original query. Final Grade: A",
' The original query is asking for shoes, so the predicted query is asking for the same thing. The original query does not specify a size, so the predicted query is not adding any additional information. The original query does not specify a price range, so the predicted query is adding additional information that is not necessary. Therefore, the predicted query is not semantically the same as the original query and is likely to produce different results. Final Grade: D',
' The original query is asking for a skirt, so the predicted query is asking for the same thing. The predicted query also adds additional parameters such as size and price range, which could help narrow down the results. However, the size parameter is not necessary for the query to be successful, and the price range is too narrow. Therefore, the predicted query is not as effective as the original query. Final Grade: C', | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-11 | ' The first part of the query is asking for a Desktop PC, which is the same as the original query. The second part of the query is asking for a size of 10, which is not relevant to the original query. The third part of the query is asking for a minimum price of 0, which is not relevant to the original query. The fourth part of the query is asking for a maximum price of null, which is not relevant to the original query. Therefore, the Predicted Query does not semantically match the original query and is not likely to produce the same answer. Final Grade: F',
' The original query is asking for cameras with a maximum price of 300. The predicted query is asking for cameras with a maximum price of 500. This means that the predicted query is likely to return more results than the original query, which may include cameras that are not within the budget range. Therefore, the predicted query is not semantically the same as the original query and does not answer the original question. Final Grade: F']
import re
from typing import List
# Parse the evaluation chain responses into a rubric
def parse_eval_results(results: List[str]) -> List[float]:
rubric = {
"A": 1.0,
"B": 0.75,
"C": 0.5,
"D": 0.25,
"F": 0
}
return [rubric[re.search(r'Final Grade: (\w+)', res).group(1)] for res in results]
parsed_results = parse_eval_results(request_eval_results)
# Collect the scores for a final evaluation table
scores['request_synthesizer'].extend(parsed_results)
Evaluate the Response Chain#
The second component translated the structured API response to a natural language response.
Evaluate this against the user’s original question. | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-12 | Evaluate this against the user’s original question.
from langchain.prompts import PromptTemplate
template = """You are trying to answer the following question by querying an API:
> Question: {question}
The API returned a response of:
> API result: {api_response}
Your response to the user: {answer}
Please evaluate the accuracy and utility of your response to the user's original question, conditioned on the information available.
Give a letter grade of either an A, B, C, D, or F, along with an explanation of why. End the evaluation with 'Final Grade: <the letter>'
> Explanation: Let's think step by step."""
prompt = PromptTemplate.from_template(template)
eval_chain = LLMChain(llm=llm, prompt=prompt, verbose=verbose)
# Extract the API responses from the chain
api_responses = [output["intermediate_steps"]["response_text"] for output in chain_outputs]
# Run the grader chain
response_eval_results = []
for question, api_response, answer in list(zip(questions, api_responses, answers)):
request_eval_results.append(eval_chain.run(question=question, api_response=api_response, answer=answer))
request_eval_results | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-13 | request_eval_results
[' The original query is asking for all iPhone models, so the "q" parameter is correct. The "max_price" parameter is also correct, as it is set to null, meaning that no maximum price is set. The predicted query adds two additional parameters, "size" and "min_price". The "size" parameter is not necessary, as it is not relevant to the question being asked. The "min_price" parameter is also not necessary, as it is not relevant to the question being asked and it is set to 0, which is the default value. Therefore, the predicted query is not semantically the same as the original query and is not likely to produce the same answer. Final Grade: D',
' The original query is asking for laptops with a maximum price of 300. The predicted query is asking for laptops with a minimum price of 0 and a maximum price of 500. This means that the predicted query is likely to return more results than the original query, as it is asking for a wider range of prices. Therefore, the predicted query is not semantically the same as the original query, and it is not likely to produce the same answer. Final Grade: F',
" The first two parameters are the same, so that's good. The third parameter is different, but it's not necessary for the query, so that's not a problem. The fourth parameter is the problem. The original query specifies a maximum price of 500, while the predicted query specifies a maximum price of null. This means that the predicted query will not limit the results to the cheapest gaming PCs, so it is not semantically the same as the original query. Final Grade: F", | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-14 | ' The original query is asking for tablets under $400, so the first two parameters are correct. The predicted query also includes the parameters "size" and "min_price", which are not necessary for the original query. The "size" parameter is not relevant to the question, and the "min_price" parameter is redundant since the original query already specifies a maximum price. Therefore, the predicted query is not semantically the same as the original query and is not likely to produce the same answer. Final Grade: D',
' The original query is asking for headphones with no maximum price, so the predicted query is not semantically the same because it has a maximum price of 500. The predicted query also has a size of 10, which is not specified in the original query. Therefore, the predicted query is not semantically the same as the original query. Final Grade: F',
" The original query is asking for the top rated laptops, so the 'size' parameter should be set to 10 to get the top 10 results. The 'min_price' parameter should be set to 0 to get results from all price ranges. The 'max_price' parameter should be set to null to get results from all price ranges. The 'q' parameter should be set to 'laptop' to get results related to laptops. All of these parameters are present in the predicted query, so it is semantically the same as the original query. Final Grade: A",
' The original query is asking for shoes, so the predicted query is asking for the same thing. The original query does not specify a size, so the predicted query is not adding any additional information. The original query does not specify a price range, so the predicted query is adding additional information that is not necessary. Therefore, the predicted query is not semantically the same as the original query and is likely to produce different results. Final Grade: D', | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-15 | ' The original query is asking for a skirt, so the predicted query is asking for the same thing. The predicted query also adds additional parameters such as size and price range, which could help narrow down the results. However, the size parameter is not necessary for the query to be successful, and the price range is too narrow. Therefore, the predicted query is not as effective as the original query. Final Grade: C',
' The first part of the query is asking for a Desktop PC, which is the same as the original query. The second part of the query is asking for a size of 10, which is not relevant to the original query. The third part of the query is asking for a minimum price of 0, which is not relevant to the original query. The fourth part of the query is asking for a maximum price of null, which is not relevant to the original query. Therefore, the Predicted Query does not semantically match the original query and is not likely to produce the same answer. Final Grade: F',
' The original query is asking for cameras with a maximum price of 300. The predicted query is asking for cameras with a maximum price of 500. This means that the predicted query is likely to return more results than the original query, which may include cameras that are not within the budget range. Therefore, the predicted query is not semantically the same as the original query and does not answer the original question. Final Grade: F',
' The user asked a question about what iPhone models are available, and the API returned a response with 10 different models. The response provided by the user accurately listed all 10 models, so the accuracy of the response is A+. The utility of the response is also A+ since the user was able to get the exact information they were looking for. Final Grade: A+', | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-16 | " The API response provided a list of laptops with their prices and attributes. The user asked if there were any budget laptops, and the response provided a list of laptops that are all priced under $500. Therefore, the response was accurate and useful in answering the user's question. Final Grade: A",
" The API response provided the name, price, and URL of the product, which is exactly what the user asked for. The response also provided additional information about the product's attributes, which is useful for the user to make an informed decision. Therefore, the response is accurate and useful. Final Grade: A",
" The API response provided a list of tablets that are under $400. The response accurately answered the user's question. Additionally, the response provided useful information such as the product name, price, and attributes. Therefore, the response was accurate and useful. Final Grade: A",
" The API response provided a list of headphones with their respective prices and attributes. The user asked for the best headphones, so the response should include the best headphones based on the criteria provided. The response provided a list of headphones that are all from the same brand (Apple) and all have the same type of headphone (True Wireless, In-Ear). This does not provide the user with enough information to make an informed decision about which headphones are the best. Therefore, the response does not accurately answer the user's question. Final Grade: F",
' The API response provided a list of laptops with their attributes, which is exactly what the user asked for. The response provided a comprehensive list of the top rated laptops, which is what the user was looking for. The response was accurate and useful, providing the user with the information they needed. Final Grade: A', | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-17 | ' The API response provided a list of shoes from both Adidas and Nike, which is exactly what the user asked for. The response also included the product name, price, and attributes for each shoe, which is useful information for the user to make an informed decision. The response also included links to the products, which is helpful for the user to purchase the shoes. Therefore, the response was accurate and useful. Final Grade: A',
" The API response provided a list of skirts that could potentially meet the user's needs. The response also included the name, price, and attributes of each skirt. This is a great start, as it provides the user with a variety of options to choose from. However, the response does not provide any images of the skirts, which would have been helpful for the user to make a decision. Additionally, the response does not provide any information about the availability of the skirts, which could be important for the user. \n\nFinal Grade: B",
' The user asked for a professional desktop PC with no budget constraints. The API response provided a list of products that fit the criteria, including the Skytech Archangel Gaming Computer PC Desktop, the CyberPowerPC Gamer Master Gaming Desktop, and the ASUS ROG Strix G10DK-RS756. The response accurately suggested these three products as they all offer powerful processors and plenty of RAM. Therefore, the response is accurate and useful. Final Grade: A',
" The API response provided a list of cameras with their prices, which is exactly what the user asked for. The response also included additional information such as features and memory cards, which is not necessary for the user's question but could be useful for further research. The response was accurate and provided the user with the information they needed. Final Grade: A"]
# Reusing the rubric from above, parse the evaluation chain responses
parsed_response_results = parse_eval_results(request_eval_results) | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-18 | parsed_response_results = parse_eval_results(request_eval_results)
# Collect the scores for a final evaluation table
scores['result_synthesizer'].extend(parsed_response_results)
# Print out Score statistics for the evaluation session
header = "{:<20}\t{:<10}\t{:<10}\t{:<10}".format("Metric", "Min", "Mean", "Max")
print(header)
for metric, metric_scores in scores.items():
mean_scores = sum(metric_scores) / len(metric_scores) if len(metric_scores) > 0 else float('nan')
row = "{:<20}\t{:<10.2f}\t{:<10.2f}\t{:<10.2f}".format(metric, min(metric_scores), mean_scores, max(metric_scores))
print(row)
Metric Min Mean Max
completed 1.00 1.00 1.00
request_synthesizer 0.00 0.23 1.00
result_synthesizer 0.00 0.55 1.00
# Re-show the examples for which the chain failed to complete
failed_examples
[]
Generating Test Datasets#
To evaluate a chain against your own endpoint, you’ll want to generate a test dataset that’s conforms to the API.
This section provides an overview of how to bootstrap the process.
First, we’ll parse the OpenAPI Spec. For this example, we’ll Speak’s OpenAPI specification.
# Load and parse the OpenAPI Spec
spec = OpenAPISpec.from_url("https://api.speak.com/openapi.yaml")
Attempting to load an OpenAPI 3.0.1 spec. This may result in degraded performance. Convert your OpenAPI spec to 3.1.* spec for better support. | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-19 | Attempting to load an OpenAPI 3.0.1 spec. This may result in degraded performance. Convert your OpenAPI spec to 3.1.* spec for better support.
# List the paths in the OpenAPI Spec
paths = sorted(spec.paths.keys())
paths
['/v1/public/openai/explain-phrase',
'/v1/public/openai/explain-task',
'/v1/public/openai/translate']
# See which HTTP Methods are available for a given path
methods = spec.get_methods_for_path('/v1/public/openai/explain-task')
methods
['post']
# Load a single endpoint operation
operation = APIOperation.from_openapi_spec(spec, '/v1/public/openai/explain-task', 'post')
# The operation can be serialized as typescript
print(operation.to_typescript())
type explainTask = (_: {
/* Description of the task that the user wants to accomplish or do. For example, "tell the waiter they messed up my order" or "compliment someone on their shirt" */
task_description?: string,
/* The foreign language that the user is learning and asking about. The value can be inferred from question - for example, if the user asks "how do i ask a girl out in mexico city", the value should be "Spanish" because of Mexico City. Always use the full name of the language (e.g. Spanish, French). */
learning_language?: string,
/* The user's native language. Infer this value from the language the user asked their question in. Always use the full name of the language (e.g. Spanish, French). */
native_language?: string,
/* A description of any additional context in the user's question that could affect the explanation - e.g. setting, scenario, situation, tone, speaking style and formality, usage notes, or any other qualifiers. */ | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-20 | additional_context?: string,
/* Full text of the user's question. */
full_query?: string,
}) => any;
# Compress the service definition to avoid leaking too much input structure to the sample data
template = """In 20 words or less, what does this service accomplish?
{spec}
Function: It's designed to """
prompt = PromptTemplate.from_template(template)
generation_chain = LLMChain(llm=llm, prompt=prompt)
purpose = generation_chain.run(spec=operation.to_typescript())
template = """Write a list of {num_to_generate} unique messages users might send to a service designed to{purpose} They must each be completely unique.
1."""
def parse_list(text: str) -> List[str]:
# Match lines starting with a number then period
# Strip leading and trailing whitespace
matches = re.findall(r'^\d+\. ', text)
return [re.sub(r'^\d+\. ', '', q).strip().strip('"') for q in text.split('\n')]
num_to_generate = 10 # How many examples to use for this test set.
prompt = PromptTemplate.from_template(template)
generation_chain = LLMChain(llm=llm, prompt=prompt)
text = generation_chain.run(purpose=purpose,
num_to_generate=num_to_generate)
# Strip preceding numeric bullets
queries = parse_list(text)
queries
["Can you explain how to say 'hello' in Spanish?",
"I need help understanding the French word for 'goodbye'.",
"Can you tell me how to say 'thank you' in German?",
"I'm trying to learn the Italian word for 'please'.",
"Can you help me with the pronunciation of 'yes' in Portuguese?",
"I'm looking for the Dutch word for 'no'.", | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-21 | "I'm looking for the Dutch word for 'no'.",
"Can you explain the meaning of 'hello' in Japanese?",
"I need help understanding the Russian word for 'thank you'.",
"Can you tell me how to say 'goodbye' in Chinese?",
"I'm trying to learn the Arabic word for 'please'."]
# Define the generation chain to get hypotheses
api_chain = OpenAPIEndpointChain.from_api_operation(
operation,
llm,
requests=Requests(),
verbose=verbose,
return_intermediate_steps=True # Return request and response text
)
predicted_outputs =[api_chain(query) for query in queries]
request_args = [output["intermediate_steps"]["request_args"] for output in predicted_outputs]
# Show the generated request
request_args
['{"task_description": "say \'hello\'", "learning_language": "Spanish", "native_language": "English", "full_query": "Can you explain how to say \'hello\' in Spanish?"}',
'{"task_description": "understanding the French word for \'goodbye\'", "learning_language": "French", "native_language": "English", "full_query": "I need help understanding the French word for \'goodbye\'."}',
'{"task_description": "say \'thank you\'", "learning_language": "German", "native_language": "English", "full_query": "Can you tell me how to say \'thank you\' in German?"}',
'{"task_description": "Learn the Italian word for \'please\'", "learning_language": "Italian", "native_language": "English", "full_query": "I\'m trying to learn the Italian word for \'please\'."}', | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-22 | '{"task_description": "Help with pronunciation of \'yes\' in Portuguese", "learning_language": "Portuguese", "native_language": "English", "full_query": "Can you help me with the pronunciation of \'yes\' in Portuguese?"}',
'{"task_description": "Find the Dutch word for \'no\'", "learning_language": "Dutch", "native_language": "English", "full_query": "I\'m looking for the Dutch word for \'no\'."}',
'{"task_description": "Explain the meaning of \'hello\' in Japanese", "learning_language": "Japanese", "native_language": "English", "full_query": "Can you explain the meaning of \'hello\' in Japanese?"}',
'{"task_description": "understanding the Russian word for \'thank you\'", "learning_language": "Russian", "native_language": "English", "full_query": "I need help understanding the Russian word for \'thank you\'."}',
'{"task_description": "say goodbye", "learning_language": "Chinese", "native_language": "English", "full_query": "Can you tell me how to say \'goodbye\' in Chinese?"}',
'{"task_description": "Learn the Arabic word for \'please\'", "learning_language": "Arabic", "native_language": "English", "full_query": "I\'m trying to learn the Arabic word for \'please\'."}']
## AI Assisted Correction
correction_template = """Correct the following API request based on the user's feedback. If the user indicates no changes are needed, output the original without making any changes.
REQUEST: {request}
User Feedback / requested changes: {user_feedback}
Finalized Request: """
prompt = PromptTemplate.from_template(correction_template)
correction_chain = LLMChain(llm=llm, prompt=prompt)
ground_truth = [] | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-23 | ground_truth = []
for query, request_arg in list(zip(queries, request_args)):
feedback = input(f"Query: {query}\nRequest: {request_arg}\nRequested changes: ")
if feedback == 'n' or feedback == 'none' or not feedback:
ground_truth.append(request_arg)
continue
resolved = correction_chain.run(request=request_arg,
user_feedback=feedback)
ground_truth.append(resolved.strip())
print("Updated request:", resolved)
Query: Can you explain how to say 'hello' in Spanish?
Request: {"task_description": "say 'hello'", "learning_language": "Spanish", "native_language": "English", "full_query": "Can you explain how to say 'hello' in Spanish?"}
Requested changes:
Query: I need help understanding the French word for 'goodbye'.
Request: {"task_description": "understanding the French word for 'goodbye'", "learning_language": "French", "native_language": "English", "full_query": "I need help understanding the French word for 'goodbye'."}
Requested changes:
Query: Can you tell me how to say 'thank you' in German?
Request: {"task_description": "say 'thank you'", "learning_language": "German", "native_language": "English", "full_query": "Can you tell me how to say 'thank you' in German?"}
Requested changes:
Query: I'm trying to learn the Italian word for 'please'.
Request: {"task_description": "Learn the Italian word for 'please'", "learning_language": "Italian", "native_language": "English", "full_query": "I'm trying to learn the Italian word for 'please'."}
Requested changes:
Query: Can you help me with the pronunciation of 'yes' in Portuguese? | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-24 | Query: Can you help me with the pronunciation of 'yes' in Portuguese?
Request: {"task_description": "Help with pronunciation of 'yes' in Portuguese", "learning_language": "Portuguese", "native_language": "English", "full_query": "Can you help me with the pronunciation of 'yes' in Portuguese?"}
Requested changes:
Query: I'm looking for the Dutch word for 'no'.
Request: {"task_description": "Find the Dutch word for 'no'", "learning_language": "Dutch", "native_language": "English", "full_query": "I'm looking for the Dutch word for 'no'."}
Requested changes:
Query: Can you explain the meaning of 'hello' in Japanese?
Request: {"task_description": "Explain the meaning of 'hello' in Japanese", "learning_language": "Japanese", "native_language": "English", "full_query": "Can you explain the meaning of 'hello' in Japanese?"}
Requested changes:
Query: I need help understanding the Russian word for 'thank you'.
Request: {"task_description": "understanding the Russian word for 'thank you'", "learning_language": "Russian", "native_language": "English", "full_query": "I need help understanding the Russian word for 'thank you'."}
Requested changes:
Query: Can you tell me how to say 'goodbye' in Chinese?
Request: {"task_description": "say goodbye", "learning_language": "Chinese", "native_language": "English", "full_query": "Can you tell me how to say 'goodbye' in Chinese?"}
Requested changes:
Query: I'm trying to learn the Arabic word for 'please'. | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-25 | Requested changes:
Query: I'm trying to learn the Arabic word for 'please'.
Request: {"task_description": "Learn the Arabic word for 'please'", "learning_language": "Arabic", "native_language": "English", "full_query": "I'm trying to learn the Arabic word for 'please'."}
Requested changes:
Now you can use the ground_truth as shown above in Evaluate the Requests Chain!
# Now you have a new ground truth set to use as shown above!
ground_truth
['{"task_description": "say \'hello\'", "learning_language": "Spanish", "native_language": "English", "full_query": "Can you explain how to say \'hello\' in Spanish?"}',
'{"task_description": "understanding the French word for \'goodbye\'", "learning_language": "French", "native_language": "English", "full_query": "I need help understanding the French word for \'goodbye\'."}',
'{"task_description": "say \'thank you\'", "learning_language": "German", "native_language": "English", "full_query": "Can you tell me how to say \'thank you\' in German?"}',
'{"task_description": "Learn the Italian word for \'please\'", "learning_language": "Italian", "native_language": "English", "full_query": "I\'m trying to learn the Italian word for \'please\'."}',
'{"task_description": "Help with pronunciation of \'yes\' in Portuguese", "learning_language": "Portuguese", "native_language": "English", "full_query": "Can you help me with the pronunciation of \'yes\' in Portuguese?"}',
'{"task_description": "Find the Dutch word for \'no\'", "learning_language": "Dutch", "native_language": "English", "full_query": "I\'m looking for the Dutch word for \'no\'."}', | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
54a638d45033-26 | '{"task_description": "Explain the meaning of \'hello\' in Japanese", "learning_language": "Japanese", "native_language": "English", "full_query": "Can you explain the meaning of \'hello\' in Japanese?"}',
'{"task_description": "understanding the Russian word for \'thank you\'", "learning_language": "Russian", "native_language": "English", "full_query": "I need help understanding the Russian word for \'thank you\'."}',
'{"task_description": "say goodbye", "learning_language": "Chinese", "native_language": "English", "full_query": "Can you tell me how to say \'goodbye\' in Chinese?"}',
'{"task_description": "Learn the Arabic word for \'please\'", "learning_language": "Arabic", "native_language": "English", "full_query": "I\'m trying to learn the Arabic word for \'please\'."}']
previous
LLM Math
next
Question Answering Benchmarking: Paul Graham Essay
Contents
Load the API Chain
Optional: Generate Input Questions and Request Ground Truth Queries
Run the API Chain
Evaluate the requests chain
Evaluate the Response Chain
Generating Test Datasets
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/evaluation/openapi_eval.html |
e757278fab64-0 | .ipynb
.pdf
Question Answering
Contents
Setup
Examples
Predictions
Evaluation
Customize Prompt
Evaluation without Ground Truth
Comparing to other evaluation metrics
Question Answering#
This notebook covers how to evaluate generic question answering problems. This is a situation where you have an example containing a question and its corresponding ground truth answer, and you want to measure how well the language model does at answering those questions.
Setup#
For demonstration purposes, we will just evaluate a simple question answering system that only evaluates the model’s internal knowledge. Please see other notebooks for examples where it evaluates how the model does at question answering over data not present in what the model was trained on.
from langchain.prompts import PromptTemplate
from langchain.chains import LLMChain
from langchain.llms import OpenAI
prompt = PromptTemplate(template="Question: {question}\nAnswer:", input_variables=["question"])
llm = OpenAI(model_name="text-davinci-003", temperature=0)
chain = LLMChain(llm=llm, prompt=prompt)
Examples#
For this purpose, we will just use two simple hardcoded examples, but see other notebooks for tips on how to get and/or generate these examples.
examples = [
{
"question": "Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now?",
"answer": "11"
},
{
"question": 'Is the following sentence plausible? "Joao Moutinho caught the screen pass in the NFC championship."',
"answer": "No"
}
]
Predictions#
We can now make and inspect the predictions for these questions.
predictions = chain.apply(examples)
predictions
[{'text': ' 11 tennis balls'}, | https://python.langchain.com/en/latest/use_cases/evaluation/question_answering.html |
e757278fab64-1 | predictions = chain.apply(examples)
predictions
[{'text': ' 11 tennis balls'},
{'text': ' No, this sentence is not plausible. Joao Moutinho is a professional soccer player, not an American football player, so it is not likely that he would be catching a screen pass in the NFC championship.'}]
Evaluation#
We can see that if we tried to just do exact match on the answer answers (11 and No) they would not match what the language model answered. However, semantically the language model is correct in both cases. In order to account for this, we can use a language model itself to evaluate the answers.
from langchain.evaluation.qa import QAEvalChain
llm = OpenAI(temperature=0)
eval_chain = QAEvalChain.from_llm(llm)
graded_outputs = eval_chain.evaluate(examples, predictions, question_key="question", prediction_key="text")
for i, eg in enumerate(examples):
print(f"Example {i}:")
print("Question: " + eg['question'])
print("Real Answer: " + eg['answer'])
print("Predicted Answer: " + predictions[i]['text'])
print("Predicted Grade: " + graded_outputs[i]['text'])
print()
Example 0:
Question: Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now?
Real Answer: 11
Predicted Answer: 11 tennis balls
Predicted Grade: CORRECT
Example 1:
Question: Is the following sentence plausible? "Joao Moutinho caught the screen pass in the NFC championship."
Real Answer: No | https://python.langchain.com/en/latest/use_cases/evaluation/question_answering.html |
e757278fab64-2 | Real Answer: No
Predicted Answer: No, this sentence is not plausible. Joao Moutinho is a professional soccer player, not an American football player, so it is not likely that he would be catching a screen pass in the NFC championship.
Predicted Grade: CORRECT
Customize Prompt#
You can also customize the prompt that is used. Here is an example prompting it using a score from 0 to 10.
The custom prompt requires 3 input variables: “query”, “answer” and “result”. Where “query” is the question, “answer” is the ground truth answer, and “result” is the predicted answer.
from langchain.prompts.prompt import PromptTemplate
_PROMPT_TEMPLATE = """You are an expert professor specialized in grading students' answers to questions.
You are grading the following question:
{query}
Here is the real answer:
{answer}
You are grading the following predicted answer:
{result}
What grade do you give from 0 to 10, where 0 is the lowest (very low similarity) and 10 is the highest (very high similarity)?
"""
PROMPT = PromptTemplate(input_variables=["query", "answer", "result"], template=_PROMPT_TEMPLATE)
evalchain = QAEvalChain.from_llm(llm=llm,prompt=PROMPT)
evalchain.evaluate(examples, predictions, question_key="question", answer_key="answer", prediction_key="text")
Evaluation without Ground Truth#
Its possible to evaluate question answering systems without ground truth. You would need a "context" input that reflects what the information the LLM uses to answer the question. This context can be obtained by any retreival system. Here’s an example of how it works:
context_examples = [
{
"question": "How old am I?", | https://python.langchain.com/en/latest/use_cases/evaluation/question_answering.html |
e757278fab64-3 | context_examples = [
{
"question": "How old am I?",
"context": "I am 30 years old. I live in New York and take the train to work everyday.",
},
{
"question": 'Who won the NFC championship game in 2023?"',
"context": "NFC Championship Game 2023: Philadelphia Eagles 31, San Francisco 49ers 7"
}
]
QA_PROMPT = "Answer the question based on the context\nContext:{context}\nQuestion:{question}\nAnswer:"
template = PromptTemplate(input_variables=["context", "question"], template=QA_PROMPT)
qa_chain = LLMChain(llm=llm, prompt=template)
predictions = qa_chain.apply(context_examples)
predictions
[{'text': 'You are 30 years old.'},
{'text': ' The Philadelphia Eagles won the NFC championship game in 2023.'}]
from langchain.evaluation.qa import ContextQAEvalChain
eval_chain = ContextQAEvalChain.from_llm(llm)
graded_outputs = eval_chain.evaluate(context_examples, predictions, question_key="question", prediction_key="text")
graded_outputs
[{'text': ' CORRECT'}, {'text': ' CORRECT'}]
Comparing to other evaluation metrics#
We can compare the evaluation results we get to other common evaluation metrics. To do this, let’s load some evaluation metrics from HuggingFace’s evaluate package.
# Some data munging to get the examples in the right format
for i, eg in enumerate(examples):
eg['id'] = str(i)
eg['answers'] = {"text": [eg['answer']], "answer_start": [0]}
predictions[i]['id'] = str(i) | https://python.langchain.com/en/latest/use_cases/evaluation/question_answering.html |
e757278fab64-4 | predictions[i]['id'] = str(i)
predictions[i]['prediction_text'] = predictions[i]['text']
for p in predictions:
del p['text']
new_examples = examples.copy()
for eg in new_examples:
del eg ['question']
del eg['answer']
from evaluate import load
squad_metric = load("squad")
results = squad_metric.compute(
references=new_examples,
predictions=predictions,
)
results
{'exact_match': 0.0, 'f1': 28.125}
previous
QA Generation
next
SQL Question Answering Benchmarking: Chinook
Contents
Setup
Examples
Predictions
Evaluation
Customize Prompt
Evaluation without Ground Truth
Comparing to other evaluation metrics
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/evaluation/question_answering.html |
3e2bdc21886b-0 | .ipynb
.pdf
SQL Question Answering Benchmarking: Chinook
Contents
Loading the data
Setting up a chain
Make a prediction
Make many predictions
Evaluate performance
SQL Question Answering Benchmarking: Chinook#
Here we go over how to benchmark performance on a question answering task over a SQL database.
It is highly reccomended that you do any evaluation/benchmarking with tracing enabled. See here for an explanation of what tracing is and how to set it up.
# Comment this out if you are NOT using tracing
import os
os.environ["LANGCHAIN_HANDLER"] = "langchain"
Loading the data#
First, let’s load the data.
from langchain.evaluation.loading import load_dataset
dataset = load_dataset("sql-qa-chinook")
Downloading and preparing dataset json/LangChainDatasets--sql-qa-chinook to /Users/harrisonchase/.cache/huggingface/datasets/LangChainDatasets___json/LangChainDatasets--sql-qa-chinook-7528565d2d992b47/0.0.0/0f7e3662623656454fcd2b650f34e886a7db4b9104504885bd462096cc7a9f51...
Dataset json downloaded and prepared to /Users/harrisonchase/.cache/huggingface/datasets/LangChainDatasets___json/LangChainDatasets--sql-qa-chinook-7528565d2d992b47/0.0.0/0f7e3662623656454fcd2b650f34e886a7db4b9104504885bd462096cc7a9f51. Subsequent calls will reuse this data.
dataset[0]
{'question': 'How many employees are there?', 'answer': '8'} | https://python.langchain.com/en/latest/use_cases/evaluation/sql_qa_benchmarking_chinook.html |
3e2bdc21886b-1 | {'question': 'How many employees are there?', 'answer': '8'}
Setting up a chain#
This uses the example Chinook database.
To set it up follow the instructions on https://database.guide/2-sample-databases-sqlite/, placing the .db file in a notebooks folder at the root of this repository.
Note that here we load a simple chain. If you want to experiment with more complex chains, or an agent, just create the chain object in a different way.
from langchain import OpenAI, SQLDatabase, SQLDatabaseChain
db = SQLDatabase.from_uri("sqlite:///../../../notebooks/Chinook.db")
llm = OpenAI(temperature=0)
Now we can create a SQL database chain.
chain = SQLDatabaseChain(llm=llm, database=db, input_key="question")
Make a prediction#
First, we can make predictions one datapoint at a time. Doing it at this level of granularity allows use to explore the outputs in detail, and also is a lot cheaper than running over multiple datapoints
chain(dataset[0])
{'question': 'How many employees are there?',
'answer': '8',
'result': ' There are 8 employees.'}
Make many predictions#
Now we can make predictions. Note that we add a try-except because this chain can sometimes error (if SQL is written incorrectly, etc)
predictions = []
predicted_dataset = []
error_dataset = []
for data in dataset:
try:
predictions.append(chain(data))
predicted_dataset.append(data)
except:
error_dataset.append(data)
Evaluate performance#
Now we can evaluate the predictions. We can use a language model to score them programatically
from langchain.evaluation.qa import QAEvalChain
llm = OpenAI(temperature=0) | https://python.langchain.com/en/latest/use_cases/evaluation/sql_qa_benchmarking_chinook.html |
3e2bdc21886b-2 | llm = OpenAI(temperature=0)
eval_chain = QAEvalChain.from_llm(llm)
graded_outputs = eval_chain.evaluate(predicted_dataset, predictions, question_key="question", prediction_key="result")
We can add in the graded output to the predictions dict and then get a count of the grades.
for i, prediction in enumerate(predictions):
prediction['grade'] = graded_outputs[i]['text']
from collections import Counter
Counter([pred['grade'] for pred in predictions])
Counter({' CORRECT': 3, ' INCORRECT': 4})
We can also filter the datapoints to the incorrect examples and look at them.
incorrect = [pred for pred in predictions if pred['grade'] == " INCORRECT"]
incorrect[0]
{'question': 'How many employees are also customers?',
'answer': 'None',
'result': ' 59 employees are also customers.',
'grade': ' INCORRECT'}
previous
Question Answering
next
Installation
Contents
Loading the data
Setting up a chain
Make a prediction
Make many predictions
Evaluate performance
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/evaluation/sql_qa_benchmarking_chinook.html |
aa39727f39b5-0 | .ipynb
.pdf
Generic Agent Evaluation
Contents
Setup
Testing the Agent
Evaluating the Agent
Generic Agent Evaluation#
Good evaluation is key for quickly iterating on your agent’s prompts and tools. Here we provide an example of how to use the TrajectoryEvalChain to evaluate your agent.
Setup#
Let’s start by defining our agent.
from langchain import Wikipedia
from langchain.chat_models import ChatOpenAI
from langchain.agents import initialize_agent, Tool
from langchain.agents import AgentType
from langchain.agents.react.base import DocstoreExplorer
from langchain.memory import ConversationBufferMemory
from langchain import LLMMathChain
from langchain.llms import OpenAI
from langchain import SerpAPIWrapper
docstore = DocstoreExplorer(Wikipedia())
math_llm = OpenAI(temperature=0)
llm_math_chain = LLMMathChain(llm=math_llm, verbose=True)
search = SerpAPIWrapper()
tools = [
Tool(
name="Search",
func=docstore.search,
description="useful for when you need to ask with search",
),
Tool(
name="Lookup",
func=docstore.lookup,
description="useful for when you need to ask with lookup",
),
Tool(
name="Calculator",
func=llm_math_chain.run,
description="useful for doing calculations",
),
Tool(
name="Search the Web (SerpAPI)",
func=search.run,
description="useful for when you need to answer questions about current events",
),
]
memory = ConversationBufferMemory(
memory_key="chat_history", return_messages=True, output_key="output"
) | https://python.langchain.com/en/latest/use_cases/evaluation/generic_agent_evaluation.html |
aa39727f39b5-1 | memory_key="chat_history", return_messages=True, output_key="output"
)
llm = ChatOpenAI(temperature=0, model_name="gpt-3.5-turbo")
agent = initialize_agent(
tools,
llm,
agent=AgentType.CHAT_CONVERSATIONAL_REACT_DESCRIPTION,
verbose=True,
memory=memory,
return_intermediate_steps=True, # This is needed for the evaluation later
)
Testing the Agent#
Now let’s try our agent out on some example queries.
query_one = "How many ping pong balls would it take to fill the entire Empire State Building?"
test_outputs_one = agent({"input": query_one}, return_only_outputs=False)
> Entering new AgentExecutor chain...
{
"action": "Search the Web (SerpAPI)",
"action_input": "How many ping pong balls would it take to fill the entire Empire State Building?"
}
Observation: 12.8 billion. The volume of the Empire State Building Googles in at around 37 million ft³. A golf ball comes in at about 2.5 in³.
Thought:{
"action": "Final Answer",
"action_input": "It would take approximately 12.8 billion ping pong balls to fill the entire Empire State Building."
}
> Finished chain.
This looks good! Let’s try it out on another query.
query_two = "If you laid the Eiffel Tower end to end, how many would you need cover the US from coast to coast?"
test_outputs_two = agent({"input": query_two}, return_only_outputs=False)
> Entering new AgentExecutor chain...
{
"action": "Calculator", | https://python.langchain.com/en/latest/use_cases/evaluation/generic_agent_evaluation.html |
aa39727f39b5-2 | > Entering new AgentExecutor chain...
{
"action": "Calculator",
"action_input": "The length of the Eiffel Tower is 324 meters. The distance from coast to coast in the US is approximately 4,828 kilometers. First, we need to convert 4,828 kilometers to meters, which gives us 4,828,000 meters. To find out how many Eiffel Towers we need, we can divide 4,828,000 by 324. This gives us approximately 14,876 Eiffel Towers."
}
> Entering new LLMMathChain chain...
The length of the Eiffel Tower is 324 meters. The distance from coast to coast in the US is approximately 4,828 kilometers. First, we need to convert 4,828 kilometers to meters, which gives us 4,828,000 meters. To find out how many Eiffel Towers we need, we can divide 4,828,000 by 324. This gives us approximately 14,876 Eiffel Towers.
```text
4828000 / 324
```
...numexpr.evaluate("4828000 / 324")...
Answer: 14901.234567901234
> Finished chain.
Observation: Answer: 14901.234567901234
Thought:{
"action": "Calculator",
"action_input": "The length of the Eiffel Tower is 324 meters. The distance from coast to coast in the US is approximately 4,828 kilometers. First, we need to convert 4,828 kilometers to meters, which gives us 4,828,000 meters. To find out how many Eiffel Towers we need, we can divide 4,828,000 by 324. This gives us approximately 14,901 Eiffel Towers."
} | https://python.langchain.com/en/latest/use_cases/evaluation/generic_agent_evaluation.html |
aa39727f39b5-3 | }
> Entering new LLMMathChain chain...
The length of the Eiffel Tower is 324 meters. The distance from coast to coast in the US is approximately 4,828 kilometers. First, we need to convert 4,828 kilometers to meters, which gives us 4,828,000 meters. To find out how many Eiffel Towers we need, we can divide 4,828,000 by 324. This gives us approximately 14,901 Eiffel Towers.
```text
4828000 / 324
```
...numexpr.evaluate("4828000 / 324")...
Answer: 14901.234567901234
> Finished chain.
Observation: Answer: 14901.234567901234
Thought:{
"action": "Final Answer",
"action_input": "If you laid the Eiffel Tower end to end, you would need approximately 14,901 Eiffel Towers to cover the US from coast to coast."
}
> Finished chain.
This doesn’t look so good. Let’s try running some evaluation.
Evaluating the Agent#
Let’s start by defining the TrajectoryEvalChain.
from langchain.evaluation.agents import TrajectoryEvalChain
# Define chain
eval_chain = TrajectoryEvalChain.from_llm(
llm=ChatOpenAI(temperature=0, model_name="gpt-4"), # Note: This must be a ChatOpenAI model
agent_tools=agent.tools,
return_reasoning=True,
)
Let’s try evaluating the first query.
question, steps, answer = test_outputs_one["input"], test_outputs_one["intermediate_steps"], test_outputs_one["output"]
evaluation = eval_chain(
inputs={"question": question, "answer": answer, "agent_trajectory": eval_chain.get_agent_trajectory(steps)}, | https://python.langchain.com/en/latest/use_cases/evaluation/generic_agent_evaluation.html |
aa39727f39b5-4 | )
print("Score from 1 to 5: ", evaluation["score"])
print("Reasoning: ", evaluation["reasoning"])
Score from 1 to 5: 1
Reasoning: First, let's evaluate the final answer. The final answer is incorrect because it uses the volume of golf balls instead of ping pong balls. The answer is not helpful.
Second, does the model use a logical sequence of tools to answer the question? The model only used one tool, which was the Search the Web (SerpAPI). It did not use the Calculator tool to calculate the correct volume of ping pong balls.
Third, does the AI language model use the tools in a helpful way? The model used the Search the Web (SerpAPI) tool, but the output was not helpful because it provided information about golf balls instead of ping pong balls.
Fourth, does the AI language model use too many steps to answer the question? The model used only one step, which is not too many. However, it should have used more steps to provide a correct answer.
Fifth, are the appropriate tools used to answer the question? The model should have used the Search tool to find the volume of the Empire State Building and the volume of a ping pong ball. Then, it should have used the Calculator tool to calculate the number of ping pong balls needed to fill the building.
Judgment: Given the incorrect final answer and the inappropriate use of tools, we give the model a score of 1.
That seems about right. Let’s try the second query.
question, steps, answer = test_outputs_two["input"], test_outputs_two["intermediate_steps"], test_outputs_two["output"]
evaluation = eval_chain(
inputs={"question": question, "answer": answer, "agent_trajectory": eval_chain.get_agent_trajectory(steps)},
) | https://python.langchain.com/en/latest/use_cases/evaluation/generic_agent_evaluation.html |
aa39727f39b5-5 | )
print("Score from 1 to 5: ", evaluation["score"])
print("Reasoning: ", evaluation["reasoning"])
Score from 1 to 5: 3
Reasoning: i. Is the final answer helpful?
Yes, the final answer is helpful as it provides an approximate number of Eiffel Towers needed to cover the US from coast to coast.
ii. Does the AI language use a logical sequence of tools to answer the question?
No, the AI language model does not use a logical sequence of tools. It directly uses the Calculator tool without first using the Search or Lookup tools to find the necessary information (length of the Eiffel Tower and distance from coast to coast in the US).
iii. Does the AI language model use the tools in a helpful way?
The AI language model uses the Calculator tool in a helpful way to perform the calculation, but it should have used the Search or Lookup tools first to find the required information.
iv. Does the AI language model use too many steps to answer the question?
No, the AI language model does not use too many steps. However, it repeats the same step twice, which is unnecessary.
v. Are the appropriate tools used to answer the question?
Not entirely. The AI language model should have used the Search or Lookup tools to find the required information before using the Calculator tool.
Given the above evaluation, the AI language model's performance can be scored as follows:
That also sounds about right. In conclusion, the TrajectoryEvalChain allows us to use GPT-4 to score both our agent’s outputs and tool use in addition to giving us the reasoning behind the evaluation.
previous
Data Augmented Question Answering
next
Using Hugging Face Datasets
Contents
Setup
Testing the Agent
Evaluating the Agent
By Harrison Chase | https://python.langchain.com/en/latest/use_cases/evaluation/generic_agent_evaluation.html |
aa39727f39b5-6 | Setup
Testing the Agent
Evaluating the Agent
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/evaluation/generic_agent_evaluation.html |
324be35fe685-0 | .ipynb
.pdf
Agent VectorDB Question Answering Benchmarking
Contents
Loading the data
Setting up a chain
Make a prediction
Make many predictions
Evaluate performance
Agent VectorDB Question Answering Benchmarking#
Here we go over how to benchmark performance on a question answering task using an agent to route between multiple vectordatabases.
It is highly reccomended that you do any evaluation/benchmarking with tracing enabled. See here for an explanation of what tracing is and how to set it up.
# Comment this out if you are NOT using tracing
import os
os.environ["LANGCHAIN_HANDLER"] = "langchain"
Loading the data#
First, let’s load the data.
from langchain.evaluation.loading import load_dataset
dataset = load_dataset("agent-vectordb-qa-sota-pg")
Found cached dataset json (/Users/qt/.cache/huggingface/datasets/LangChainDatasets___json/LangChainDatasets--agent-vectordb-qa-sota-pg-d3ae24016b514f92/0.0.0/fe5dd6ea2639a6df622901539cb550cf8797e5a6b2dd7af1cf934bed8e233e6e)
100%|██████████| 1/1 [00:00<00:00, 414.42it/s]
dataset[0]
{'question': 'What is the purpose of the NATO Alliance?',
'answer': 'The purpose of the NATO Alliance is to secure peace and stability in Europe after World War 2.',
'steps': [{'tool': 'State of Union QA System', 'tool_input': None},
{'tool': None, 'tool_input': 'What is the purpose of the NATO Alliance?'}]}
dataset[-1] | https://python.langchain.com/en/latest/use_cases/evaluation/agent_vectordb_sota_pg.html |
324be35fe685-1 | dataset[-1]
{'question': 'What is the purpose of YC?',
'answer': 'The purpose of YC is to cause startups to be founded that would not otherwise have existed.',
'steps': [{'tool': 'Paul Graham QA System', 'tool_input': None},
{'tool': None, 'tool_input': 'What is the purpose of YC?'}]}
Setting up a chain#
Now we need to create some pipelines for doing question answering. Step one in that is creating indexes over the data in question.
from langchain.document_loaders import TextLoader
loader = TextLoader("../../modules/state_of_the_union.txt")
from langchain.indexes import VectorstoreIndexCreator
vectorstore_sota = VectorstoreIndexCreator(vectorstore_kwargs={"collection_name":"sota"}).from_loaders([loader]).vectorstore
Using embedded DuckDB without persistence: data will be transient
Now we can create a question answering chain.
from langchain.chains import RetrievalQA
from langchain.llms import OpenAI
chain_sota = RetrievalQA.from_chain_type(llm=OpenAI(temperature=0), chain_type="stuff", retriever=vectorstore_sota.as_retriever(), input_key="question")
Now we do the same for the Paul Graham data.
loader = TextLoader("../../modules/paul_graham_essay.txt")
vectorstore_pg = VectorstoreIndexCreator(vectorstore_kwargs={"collection_name":"paul_graham"}).from_loaders([loader]).vectorstore
Using embedded DuckDB without persistence: data will be transient
chain_pg = RetrievalQA.from_chain_type(llm=OpenAI(temperature=0), chain_type="stuff", retriever=vectorstore_pg.as_retriever(), input_key="question")
We can now set up an agent to route between them.
from langchain.agents import initialize_agent, Tool | https://python.langchain.com/en/latest/use_cases/evaluation/agent_vectordb_sota_pg.html |
324be35fe685-2 | from langchain.agents import initialize_agent, Tool
from langchain.agents import AgentType
tools = [
Tool(
name = "State of Union QA System",
func=chain_sota.run,
description="useful for when you need to answer questions about the most recent state of the union address. Input should be a fully formed question."
),
Tool(
name = "Paul Graham System",
func=chain_pg.run,
description="useful for when you need to answer questions about Paul Graham. Input should be a fully formed question."
),
]
agent = initialize_agent(tools, OpenAI(temperature=0), agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION, max_iterations=4)
Make a prediction#
First, we can make predictions one datapoint at a time. Doing it at this level of granularity allows use to explore the outputs in detail, and also is a lot cheaper than running over multiple datapoints
agent.run(dataset[0]['question'])
'The purpose of the NATO Alliance is to secure peace and stability in Europe after World War 2.'
Make many predictions#
Now we can make predictions
predictions = []
predicted_dataset = []
error_dataset = []
for data in dataset:
new_data = {"input": data["question"], "answer": data["answer"]}
try:
predictions.append(agent(new_data))
predicted_dataset.append(new_data)
except Exception:
error_dataset.append(new_data)
Evaluate performance#
Now we can evaluate the predictions. The first thing we can do is look at them by eye.
predictions[0]
{'input': 'What is the purpose of the NATO Alliance?',
'answer': 'The purpose of the NATO Alliance is to secure peace and stability in Europe after World War 2.', | https://python.langchain.com/en/latest/use_cases/evaluation/agent_vectordb_sota_pg.html |
324be35fe685-3 | 'output': 'The purpose of the NATO Alliance is to secure peace and stability in Europe after World War 2.'}
Next, we can use a language model to score them programatically
from langchain.evaluation.qa import QAEvalChain
llm = OpenAI(temperature=0)
eval_chain = QAEvalChain.from_llm(llm)
graded_outputs = eval_chain.evaluate(predicted_dataset, predictions, question_key="input", prediction_key="output")
We can add in the graded output to the predictions dict and then get a count of the grades.
for i, prediction in enumerate(predictions):
prediction['grade'] = graded_outputs[i]['text']
from collections import Counter
Counter([pred['grade'] for pred in predictions])
Counter({' CORRECT': 28, ' INCORRECT': 5})
We can also filter the datapoints to the incorrect examples and look at them.
incorrect = [pred for pred in predictions if pred['grade'] == " INCORRECT"]
incorrect[0]
{'input': 'What are the four common sense steps that the author suggests to move forward safely?',
'answer': 'The four common sense steps suggested by the author to move forward safely are: stay protected with vaccines and treatments, prepare for new variants, end the shutdown of schools and businesses, and stay vigilant.',
'output': 'The four common sense steps suggested in the most recent State of the Union address are: cutting the cost of prescription drugs, providing a pathway to citizenship for Dreamers, revising laws so businesses have the workers they need and families don’t wait decades to reunite, and protecting access to health care and preserving a woman’s right to choose.',
'grade': ' INCORRECT'}
previous
Agent Benchmarking: Search + Calculator
next
Benchmarking Template
Contents
Loading the data
Setting up a chain
Make a prediction | https://python.langchain.com/en/latest/use_cases/evaluation/agent_vectordb_sota_pg.html |
324be35fe685-4 | Benchmarking Template
Contents
Loading the data
Setting up a chain
Make a prediction
Make many predictions
Evaluate performance
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/evaluation/agent_vectordb_sota_pg.html |
ea436f022c54-0 | .ipynb
.pdf
Question Answering Benchmarking: Paul Graham Essay
Contents
Loading the data
Setting up a chain
Make a prediction
Make many predictions
Evaluate performance
Question Answering Benchmarking: Paul Graham Essay#
Here we go over how to benchmark performance on a question answering task over a Paul Graham essay.
It is highly reccomended that you do any evaluation/benchmarking with tracing enabled. See here for an explanation of what tracing is and how to set it up.
# Comment this out if you are NOT using tracing
import os
os.environ["LANGCHAIN_HANDLER"] = "langchain"
Loading the data#
First, let’s load the data.
from langchain.evaluation.loading import load_dataset
dataset = load_dataset("question-answering-paul-graham")
Found cached dataset json (/Users/harrisonchase/.cache/huggingface/datasets/LangChainDatasets___json/LangChainDatasets--question-answering-paul-graham-76e8f711e038d742/0.0.0/0f7e3662623656454fcd2b650f34e886a7db4b9104504885bd462096cc7a9f51)
Setting up a chain#
Now we need to create some pipelines for doing question answering. Step one in that is creating an index over the data in question.
from langchain.document_loaders import TextLoader
loader = TextLoader("../../modules/paul_graham_essay.txt")
from langchain.indexes import VectorstoreIndexCreator
vectorstore = VectorstoreIndexCreator().from_loaders([loader]).vectorstore
Running Chroma using direct local API.
Using DuckDB in-memory for database. Data will be transient.
Now we can create a question answering chain.
from langchain.chains import RetrievalQA | https://python.langchain.com/en/latest/use_cases/evaluation/qa_benchmarking_pg.html |
ea436f022c54-1 | Now we can create a question answering chain.
from langchain.chains import RetrievalQA
from langchain.llms import OpenAI
chain = RetrievalQA.from_chain_type(llm=OpenAI(), chain_type="stuff", retriever=vectorstore.as_retriever(), input_key="question")
Make a prediction#
First, we can make predictions one datapoint at a time. Doing it at this level of granularity allows use to explore the outputs in detail, and also is a lot cheaper than running over multiple datapoints
chain(dataset[0])
{'question': 'What were the two main things the author worked on before college?',
'answer': 'The two main things the author worked on before college were writing and programming.',
'result': ' Writing and programming.'}
Make many predictions#
Now we can make predictions
predictions = chain.apply(dataset)
Evaluate performance#
Now we can evaluate the predictions. The first thing we can do is look at them by eye.
predictions[0]
{'question': 'What were the two main things the author worked on before college?',
'answer': 'The two main things the author worked on before college were writing and programming.',
'result': ' Writing and programming.'}
Next, we can use a language model to score them programatically
from langchain.evaluation.qa import QAEvalChain
llm = OpenAI(temperature=0)
eval_chain = QAEvalChain.from_llm(llm)
graded_outputs = eval_chain.evaluate(dataset, predictions, question_key="question", prediction_key="result")
We can add in the graded output to the predictions dict and then get a count of the grades.
for i, prediction in enumerate(predictions):
prediction['grade'] = graded_outputs[i]['text']
from collections import Counter
Counter([pred['grade'] for pred in predictions]) | https://python.langchain.com/en/latest/use_cases/evaluation/qa_benchmarking_pg.html |
ea436f022c54-2 | from collections import Counter
Counter([pred['grade'] for pred in predictions])
Counter({' CORRECT': 12, ' INCORRECT': 10})
We can also filter the datapoints to the incorrect examples and look at them.
incorrect = [pred for pred in predictions if pred['grade'] == " INCORRECT"]
incorrect[0]
{'question': 'What did the author write their dissertation on?',
'answer': 'The author wrote their dissertation on applications of continuations.',
'result': ' The author does not mention what their dissertation was on, so it is not known.',
'grade': ' INCORRECT'}
previous
Evaluating an OpenAPI Chain
next
Question Answering Benchmarking: State of the Union Address
Contents
Loading the data
Setting up a chain
Make a prediction
Make many predictions
Evaluate performance
By Harrison Chase
© Copyright 2023, Harrison Chase.
Last updated on May 02, 2023. | https://python.langchain.com/en/latest/use_cases/evaluation/qa_benchmarking_pg.html |
6bdfc7a36755-0 | .ipynb
.pdf
Analysis of Twitter the-algorithm source code with LangChain, GPT4 and Deep Lake
Contents
1. Index the code base (optional)
2. Question Answering on Twitter algorithm codebase
Analysis of Twitter the-algorithm source code with LangChain, GPT4 and Deep Lake#
In this tutorial, we are going to use Langchain + Deep Lake with GPT4 to analyze the code base of the twitter algorithm.
!python3 -m pip install --upgrade langchain deeplake openai tiktoken
Define OpenAI embeddings, Deep Lake multi-modal vector store api and authenticate. For full documentation of Deep Lake please follow docs and API reference.
Authenticate into Deep Lake if you want to create your own dataset and publish it. You can get an API key from the platform
import os
import getpass
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.vectorstores import DeepLake
os.environ['OPENAI_API_KEY'] = getpass.getpass('OpenAI API Key:')
os.environ['ACTIVELOOP_TOKEN'] = getpass.getpass('Activeloop Token:')
embeddings = OpenAIEmbeddings(disallowed_special=())
disallowed_special=() is required to avoid Exception: 'utf-8' codec can't decode byte 0xff in position 0: invalid start byte from tiktoken for some repositories
1. Index the code base (optional)#
You can directly skip this part and directly jump into using already indexed dataset. To begin with, first we will clone the repository, then parse and chunk the code base and use OpenAI indexing.
!git clone https://github.com/twitter/the-algorithm # replace any repository of your choice
Load all files inside the repository
import os
from langchain.document_loaders import TextLoader
root_dir = './the-algorithm'
docs = [] | https://python.langchain.com/en/latest/use_cases/code/twitter-the-algorithm-analysis-deeplake.html |
6bdfc7a36755-1 | root_dir = './the-algorithm'
docs = []
for dirpath, dirnames, filenames in os.walk(root_dir):
for file in filenames:
try:
loader = TextLoader(os.path.join(dirpath, file), encoding='utf-8')
docs.extend(loader.load_and_split())
except Exception as e:
pass
Then, chunk the files
from langchain.text_splitter import CharacterTextSplitter
text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
texts = text_splitter.split_documents(docs)
Execute the indexing. This will take about ~4 mins to compute embeddings and upload to Activeloop. You can then publish the dataset to be public.
username = "davitbun" # replace with your username from app.activeloop.ai
db = DeepLake(dataset_path=f"hub://{username}/twitter-algorithm", embedding_function=embeddings, public=True) #dataset would be publicly available
db.add_documents(texts)
2. Question Answering on Twitter algorithm codebase#
First load the dataset, construct the retriever, then construct the Conversational Chain
db = DeepLake(dataset_path="hub://davitbun/twitter-algorithm", read_only=True, embedding_function=embeddings)
retriever = db.as_retriever()
retriever.search_kwargs['distance_metric'] = 'cos'
retriever.search_kwargs['fetch_k'] = 100
retriever.search_kwargs['maximal_marginal_relevance'] = True
retriever.search_kwargs['k'] = 10
You can also specify user defined functions using Deep Lake filters
def filter(x):
# filter based on source code
if 'com.google' in x['text'].data()['value']:
return False
# filter based on path e.g. extension | https://python.langchain.com/en/latest/use_cases/code/twitter-the-algorithm-analysis-deeplake.html |
6bdfc7a36755-2 | return False
# filter based on path e.g. extension
metadata = x['metadata'].data()['value']
return 'scala' in metadata['source'] or 'py' in metadata['source']
### turn on below for custom filtering
# retriever.search_kwargs['filter'] = filter
from langchain.chat_models import ChatOpenAI
from langchain.chains import ConversationalRetrievalChain
model = ChatOpenAI(model='gpt-3.5-turbo') # switch to 'gpt-4'
qa = ConversationalRetrievalChain.from_llm(model,retriever=retriever)
questions = [
"What does favCountParams do?",
"is it Likes + Bookmarks, or not clear from the code?",
"What are the major negative modifiers that lower your linear ranking parameters?",
"How do you get assigned to SimClusters?",
"What is needed to migrate from one SimClusters to another SimClusters?",
"How much do I get boosted within my cluster?",
"How does Heavy ranker work. what are it’s main inputs?",
"How can one influence Heavy ranker?",
"why threads and long tweets do so well on the platform?",
"Are thread and long tweet creators building a following that reacts to only threads?",
"Do you need to follow different strategies to get most followers vs to get most likes and bookmarks per tweet?",
"Content meta data and how it impacts virality (e.g. ALT in images).",
"What are some unexpected fingerprints for spam factors?",
"Is there any difference between company verified checkmarks and blue verified individual checkmarks?",
]
chat_history = []
for question in questions:
result = qa({"question": question, "chat_history": chat_history}) | https://python.langchain.com/en/latest/use_cases/code/twitter-the-algorithm-analysis-deeplake.html |
6bdfc7a36755-3 | result = qa({"question": question, "chat_history": chat_history})
chat_history.append((question, result['answer']))
print(f"-> **Question**: {question} \n")
print(f"**Answer**: {result['answer']} \n")
-> Question: What does favCountParams do?
Answer: favCountParams is an optional ThriftLinearFeatureRankingParams instance that represents the parameters related to the “favorite count” feature in the ranking process. It is used to control the weight of the favorite count feature while ranking tweets. The favorite count is the number of times a tweet has been marked as a favorite by users, and it is considered an important signal in the ranking of tweets. By using favCountParams, the system can adjust the importance of the favorite count while calculating the final ranking score of a tweet.
-> Question: is it Likes + Bookmarks, or not clear from the code?
Answer: From the provided code, it is not clear if the favorite count metric is determined by the sum of likes and bookmarks. The favorite count is mentioned in the code, but there is no explicit reference to how it is calculated in terms of likes and bookmarks.
-> Question: What are the major negative modifiers that lower your linear ranking parameters?
Answer: In the given code, major negative modifiers that lower the linear ranking parameters are:
scoringData.querySpecificScore: This score adjustment is based on the query-specific information. If its value is negative, it will lower the linear ranking parameters.
scoringData.authorSpecificScore: This score adjustment is based on the author-specific information. If its value is negative, it will also lower the linear ranking parameters.
Please note that I cannot provide more information on the exact calculations of these negative modifiers, as the code for their determination is not provided.
-> Question: How do you get assigned to SimClusters? | https://python.langchain.com/en/latest/use_cases/code/twitter-the-algorithm-analysis-deeplake.html |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.