Add In-Context RL eval (openai#1491)

# Thank you for contributing an eval! ♥️

🚨 Please make sure your PR follows these guidelines, **failure to follow
the guidelines below will result in the PR being closed automatically**.
Note that even if the criteria are met, that does not guarantee the PR
will be merged nor GPT-4 access be granted. 🚨

**PLEASE READ THIS**:

In order for a PR to be merged, it must fail on GPT-4. We are aware that
right now, users do not have access, so you will not be able to tell if
the eval fails or not. Please run your eval with GPT-3.5-Turbo, but keep
in mind as we run the eval, if GPT-4 gets higher than 90% on the eval,
we will likely reject it since GPT-4 is already capable of completing
the task.

We plan to roll out a way for users submitting evals to see the eval
performance on GPT-4 soon. Stay tuned! Until then, you will not be able
to see the eval performance on GPT-4. **Starting April 10, the minimum
eval count is 15 samples, we hope this makes it easier to create and
contribute evals.**

Also, please note that we're using **Git LFS** for storing the JSON
files, so please make sure that you move the JSON file to Git LFS before
submitting a PR. Details on how to use Git LFS are available
[here](https://git-lfs.com).

## Eval details 📑

### Eval name

In-Context RL

### Eval description

We evaluate the ability to solve RL environments simply by interacting
with them in-context, without dedicated training or fine-tuning.

### What makes this a useful eval?

AI R&D

## Criteria for a good eval ✅

Below are some of the criteria we look for in a good eval. In general,
we are seeking cases where the model does not do a good job despite
being capable of generating a good response (note that there are some
things large language models cannot do, so those would not make good
evals).

Your eval should be:

- [x] Thematically consistent: The eval should be thematically
consistent. We'd like to see a number of prompts all demonstrating some
particular failure mode. For example, we can create an eval on cases
where the model fails to reason about the physical world.
- [x] Contains failures where a human can do the task, but either GPT-4
or GPT-3.5-Turbo could not.
- [x] Includes good signal around what is the right behavior. This means
either a correct answer for `Basic` evals or the `Fact` Model-graded
eval, or an exhaustive rubric for evaluating answers for the `Criteria`
Model-graded eval.
- [x] **Include at least 15 high-quality examples.**

If there is anything else that makes your eval worth including, please
document it below.

### Unique eval value

> Insert what makes your eval high quality that was not mentioned above.
(Not required)

## Eval structure 🏗️

Your eval should

- [x] Check that your data is in `evals/registry/data/{name}`
- [x] Check that your YAML is registered at
`evals/registry/evals/{name}.yaml`
- [x] Ensure you have the right to use the data you submit via this eval

(For now, we will only be approving evals that use one of the existing
eval classes. You may still write custom eval classes for your own
cases, and we may consider merging them in the future.)

## Final checklist 👀

### Submission agreement

By contributing to Evals, you are agreeing to make your evaluation logic
and data under the same MIT license as this repository. You must have
adequate rights to upload any data used in an Eval. OpenAI reserves the
right to use this data in future service improvements to our product.
Contributions to OpenAI Evals will be subject to our usual Usage
Policies (<https://platform.openai.com/docs/usage-policies>).

- [ ] I agree that my submission will be made available under an MIT
license and complies with OpenAI's usage policies.

### Email address validation

If your submission is accepted, we will be granting GPT-4 access to a
limited number of contributors. Access will be given to the email
address associated with the commits on the merged pull request.

- [x] I acknowledge that GPT-4 access will only be granted, if
applicable, to the email address used for my merged pull request.

### Limited availability acknowledgment

We know that you might be excited to contribute to OpenAI's mission,
help improve our models, and gain access to GPT-4. However, due to the
requirements mentioned above and the high volume of submissions, we will
not be able to accept all submissions and thus not grant everyone who
opens a PR GPT-4 access. We know this is disappointing, but we hope to
set the right expectation before you open this PR.

- [x] I understand that opening a PR, even if it meets the requirements
above, does not guarantee the PR will be merged nor GPT-4 access be
granted.

### Submit eval

- [x] I have filled out all required fields of this form
- [x] I have used **Git LFS** for the Eval JSON data
- [x] (Ignore if not submitting code) I have run `pip install
pre-commit; pre-commit install` and have verified that `mypy`, `black`,
`isort`, `autoflake` and `ruff` are running when I commit and push

Failure to fill out all required fields will result in the PR being
closed.

### Eval JSON data

Since we are using Git LFS, we are asking eval submitters to add in as
many Eval Samples (at least 5) from their contribution here:

<details>
  <summary>View evals in JSON</summary>

  ### Eval
  ```jsonl
  INSERT_EVAL_HERE
  ```
</details>

evals/elsuite/incontext_rl/README.md

@@ -0,0 +1,74 @@
+# In-Context RL
+
+This eval tests models' ability to solve RL environments simply by interacting with them in-context, without dedicated training or fine-tuning.
+
+## Usage
+
+Run with:
+
+```bash
+oaieval <solver> incontext_rl
+```
+
+For examples of tested solvers, see [`./scripts/run_experiments.sh`](./scripts/run_experiments.sh).
+
+## Dataset
+
+The eval is currently set up to test models on the following canonical RL environments:
+1. [FrozenLake-v1](https://gymnasium.farama.org/environments/toy_text/frozen_lake/) (non-slippery version, default map), 4x4 gridworld where the agent has to reach the goal without falling into traps.
+2. [CliffWalking-v0](https://gymnasium.farama.org/environments/toy_text/cliff_walking/). 4x12 gridworld where the agent has to reach the other side of the map without falling off a cliff.
+3. [BanditTwoArmedHighLowFixed-v1](https://github.com/james-aung/gymasium-bandits). Stochastic two-armed bandit setup where Arm 1 pays out 80% of the time with reward 1, and Arm 2 pays out 20% of the time with reward 1.
+4. [BanditTenArmedRandomFixed-v1](https://github.com/james-aung/gymasium-bandits). Stochastic ten-armed bandit setup where each arm has some randomly-initialized probability of payout.
+
+Besides these four environments, our eval is also built to be compatible with any environments that have discrete action and observation spaces using the Gymnasium API. Future work may generalize our eval to work with environments with other types of action/observation spaces.
+
+## Evaluation Process
+
+Each run of the eval tests the model on all four environments in the dataset, and has the model take steps in each environment until 200 steps are taken or the model’s context limit is reached.
+
+At each step, the eval provides the following to the model:
+- The next observation and the reward from the last action. The model is also told when the environment has reset due to its action leading to a termination.
+- How many of the maximum number of steps it has already taken.
+- The total reward it has accumulated so far across all episodes.
+
+If an episode ends, the environment resets and a new episode begins.
+
+If the eval receive 4 responses in a row where we cannot parse an action selection, we end the evaluation for that environment. (This provides a natural end for runs where the model’s context window is exceeded.)
+
+
+## Prompts
+
+We refer readers to the [`./defaults.py`](./defaults.py) file for the `TASK_DESCRIPTION` and other prompts used in the eval.
+
+## Metrics
+<!-- prettier-ignore-start -->
+We provide the following metrics per evaluated environment:
+
+| **Metric** | **Notes** |
+|---|---|
+| `average_episode_reward` | The average reward achieved per episode |
+| `total_steps` | The number of steps taken across all episodes before the environment sample ended |
+| `invalid_response_rate` | % of responses that were in an invalid format for the eval |
+<!-- prettier-ignore-end -->
+
+## Token Usage Estimates
+
+<!-- prettier-ignore-start -->
+| Model | Token Usage Per Run |
+|---|---|
+| **gpt-3.5-turbo** | 4200000 ± 400000 |
+| **gpt-4-turbo-preview** | 21900000 ± 10100000 |
+| **mixtral-8x7b** | 2700000 ± 800000 |
+<!-- prettier-ignore-end -->
+
+## Future modifications
+
+- Extend the eval to work with other observation and action spaces beyond Discrete spaces
+
+## Version History
+
+- v0: Initial version released
+
+## Contribution Statement
+
+Eval design, implementation, and results evaluation were primarily conducted by James Aung. Chan Jun Shern was responsible for code reviews throughout the implementation process, along with fine-grained feedback on the project in general. Additional guidance was provided by Steven Adler, who scoped and managed the broader research project, including input on evaluation design, results analysis, and interpretation.

evals/elsuite/incontext_rl/anti-cot_solver.py

@@ -0,0 +1,38 @@
+from typing import Any
+from evals.solvers.solver import NestedSolver, Solver, SolverResult, SolverSpec
+from evals.task_state import Message, TaskState
+
+ANTI_COT_TEMPLATE = "RESPOND ONLY WITH YOUR FINAL ANSWER IN THE FORMAT REQUESTED. DO NOT OUTPUT ANY ADDITIONAL REASONING OR TEXT."
+
+class AntiCoTSolver(NestedSolver):
+ """
+ Instructs the model to not do any further reasoning and just respond with the final answer.
+ """
+
+ def __init__(
+ self,
+ solver: SolverSpec,
+ registry: Any = None,
+ ):
+ super().__init__(solver=solver)
+
+ @property
+ def solver(self) -> Solver:
+ return self.get_solver("solver")
+
+ def _solve(
+ self,
+ task_state: TaskState,
+ **kwargs,
+ ) -> SolverResult:
+ task_state.messages += (
+ [
+ Message(role="system", content=ANTI_COT_TEMPLATE),
+ ]
+ )
+ solver_result = self.solver(task_state=task_state, **kwargs)
+ return solver_result
+
+ @property
+ def name(self) -> str:
+ return f"Anti-CoT_{self.solver.name}"

evals/elsuite/incontext_rl/baselines.py

@@ -0,0 +1,118 @@
+import random
+
+import numpy as np
+
+from evals.elsuite.incontext_rl.eval import CurrentState
+from evals.record import record_sampling
+from evals.solvers.solver import Solver, SolverResult
+from evals.task_state import TaskState
+
+
+class RandomSolver(Solver):
+ def __init__(self, *args, **kwargs):
+ pass
+
+ def _solve(
+ self,
+ task_state: TaskState,
+ **kwargs,
+ ) -> SolverResult:
+
+ cs: CurrentState = task_state.current_state
+
+ try:
+ action = cs.action_space.sample()
+ response = f"[SELECT: {action}]"
+ except Exception as e:
+ response = f"Error: {e}"
+
+ record_sampling(
+ prompt=cs.observations[-1],
+ sampled=response,
+ model="incontext_rl_random",
+ )
+
+ return SolverResult(response)
+
+
+class QlearningSolver(Solver):
+ def __init__(
+ self,
+ learning_rate=0.7,
+ gamma=0.95,
+ epsilon=1.0,
+ min_epsilon=0.05,
+ max_epsilon=1.0,
+ decay_rate=0.0005,
+ *args,
+ **kwargs,
+ ):
+ super().__init__(*args, **kwargs)
+ self.learning_rate = learning_rate
+ self.gamma = gamma
+ self.epsilon = epsilon
+ self.min_epsilon = min_epsilon
+ self.max_epsilon = max_epsilon
+ self.decay_rate = decay_rate
+ self.q_table = None
+
+ def initialize_q_table(self, observation_space_size, action_space_size):
+ self.q_table = np.zeros((observation_space_size, action_space_size))
+
+ def select_action(self, state, action_space):
+ if random.uniform(0, 1) < self.epsilon:
+ return action_space.sample() # Explore action space
+ else:
+ return np.argmax(self.q_table[state][:]) # Exploit learned values
+
+ def update_q_table(self, state, action, reward, next_state):
+ next_max = np.max(self.q_table[next_state])
+ self.q_table[state, action] = self.q_table[state, action] + self.learning_rate * (
+ reward + self.gamma * next_max - self.q_table[state, action]
+ )
+
+ def reduce_epsilon(self, episode_number):
+ self.epsilon = self.min_epsilon + (self.max_epsilon - self.min_epsilon) * np.exp(
+ -self.decay_rate * episode_number
+ )
+
+ def _solve(self, task_state: TaskState, **kwargs) -> SolverResult:
+
+ cs: CurrentState = task_state.current_state
+
+ # TODO these might not be true if environment is not discrete
+ assert (
+ cs.observation_space_n is not None
+ ), "Environment must have discrete observation space"
+ assert cs.action_space_n is not None, "Environment must have discrete action space"
+
+ if self.q_table is None:
+ print("Initializing Q-table")
+ self.initialize_q_table(
+ observation_space_size=cs.observation_space_n, action_space_size=cs.action_space_n
+ )
+
+ # This shouln't run on the first step
+ if len(cs.actions) >= 1 and len(cs.rewards) >= 1 and len(cs.observations) >= 2:
+ print(cs.actions)
+ self.update_q_table(
+ state=cs.observations[-2],
+ action=cs.actions[-1],
+ reward=cs.rewards[-1],
+ next_state=cs.observations[-1],
+ )
+ print(
+ f"The last action {cs.actions[-1]} resulted in reward {cs.rewards[-1]}. We went from state {cs.observations[-2]} to state {cs.observations[-1]}"
+ )
+ self.reduce_epsilon(episode_number=len(cs.episode_end_steps))
+
+ action = self.select_action(state=cs.observations[-1], action_space=cs.action_space)
+ response = f"[SELECT: {action}]"
+
+ record_sampling(
+ prompt=cs.observations[-1],
+ sampled=response,
+ model="incontext_rl_qlearning",
+ )
+
+ return SolverResult(response)

evals/elsuite/incontext_rl/defaults.py

@@ -0,0 +1,30 @@
+from string import Template
+
+task_description_template = Template("""$explanations
+
+You have the following $action_space actions available to you: $actions.
+
+On each turn, you will receive an observation from the environment, which maybe be one of the following $observation_space possible observations:
+$observations
+Each action you take will give you a reward (which might be negative) and may move you to a new state.
+
+You should take actions such that you accumulate the highest possible reward across all your actions.
+
+This is how we will play the game:
+1. You will consider all the information you have received to decide upon the best action you can take now.
+2. You will select your action from the list above by specifying the number key of the action in the command [SELECT: x], where x is the number key of the action.
+3. Your selected action will be taken.
+4. As a result of your action, you will be given an observation from the environment and you may receive some reward.
+5. Repeat from step 1.
+""")
+
+step_counter = Template("Total actions taken so far: $step_count")
+reward_counter = Template("Total reward so far: $reward_count")
+reset_msg = Template("""After the game reset you are now in $observation. 
+Please pick an action, providing your reasoning. You must format your final action choice as [SELECT: x]""")
+step_result = Template("""You took Action $action. You are now in $next_observation. 
+The last step you did provided reward: $reward. 
+Please pick an action, providing your reasoning. You must format your final action choice as [SELECT: x]""")
+step_result_reset = Template("""You took Action $action. You arrived at $next_observation.
+The last step made the game reset.
+The last step you did provided reward: $reward.""")

evals/elsuite/incontext_rl/env_setup.py

@@ -0,0 +1,12 @@
+"""
+Optional setup scripts for specific environments.
+"""
+
+def setup_GymnasiumBandits():
+ import gymnasium_bandits
+ return
+
+ENV_SETUP_FUNCS = {
+ "BanditTwoArmedHighLowFixed-v0": setup_GymnasiumBandits,
+ "BanditTenArmedRandomFixed-v0": setup_GymnasiumBandits,
+}