Track the Stat Eval (#1489)

@JunShern will review this

# 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

Track the stat

### Eval description

This eval measures how well models can implicitly keep track of task
state, by asking models to compute the rolling median or the rolling
mode over a sequence of integers.

### What makes this a useful eval?

[Insert why this eval is worth including and any additional context]

## 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>).

- [x] 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/track_the_stat/README.md

@@ -0,0 +1,134 @@
+# Track the Stat
+
+This eval measures how well models can implicitly keep track of task state, by
+asking models to compute the rolling median or the rolling mode over a sequence
+of integers.
+
+## Usage
+
+Run with:
+
+```bash
+oaieval <solver> track_the_stat
+```
+
+We have found that `generation/direct/gpt-4-0125-preview` works well on this
+eval. For more examples of tested solvers, see
+[`./scripts/run_experiments.sh`](./scripts/run_experiments.sh).
+
+## Evaluation Process
+
+The evaluation process is as follows for a given sample from our dataset:
+
+1. The `TASK_DESCRIPTION` prompt is shown to the solver.
+2. The sample contains an integer to use as a seed for a random number
+   generator.
+3. The random number generator generates 300 random integers between 0 and 100,
+   with replacement.
+4. The integers are shown one by one to the solver.
+5. At each turn (i.e., after each integer is shown), the solver needs to respond
+   with the current rolling median or the current rolling mode of the integers
+   seen so far.
+6. The solver's response is parsed and compared to the correct rolling median or
+   rolling mode.
+7. If the solver's response is incorrect or a violation is raised (answered in
+   the incorrect format), the evaluation stops and we measure how many turns the
+   solver lasted for. If the solver's response is correct, we move on to the
+   next integer.
+
+## Prompts
+
+We refer readers to the [`./prompts/`](./prompts/) folder for the
+`TASK_DESCRIPTION` used in the eval.
+
+## Metrics
+
+Below are the metrics returned by the eval:
+
+<!-- prettier-ignore-start -->
+| **Metric**        	| **Notes**                                                                                                                                  	|
+|-------------------	|--------------------------------------------------------------------------------------------------------------------------------------------	|
+| avg_max_length    	| The maximum sequence length the model can handle before failing, averaged across the samples. Higher is better. Best possible is 300.      	|
+| stddev_max_length 	| The standard deviation on the above.                                                                                                       	|
+| median_max_length 	| The median of the maximum sequence length the model can handle before failing, across the samples. Higher is better. Best possible is 300. 	|
+| max_max_length    	| The maximum sequence length the model handled before failing across all samples.                                                           	|
+| min_max_length    	| The minimum sequence length the model handled before failing across all samples.                                                           	|
+| violation_rate    	| how often the model responds in an invalid format. i.e. not using the `[<task>: <answer>]` format.                                         	|
+<!-- prettier-ignore-end -->
+
+## Variants
+
+The eval has two variants: median and mode. In the median variant, the solver
+needs to track the rolling median. In the mode variant, the solver needs to
+track the rolling mode.
+
+```bash
+oaieval <solver> track_the_stat.<median/mode>
+```
+
+## Custom Solvers
+
+We implement 3 custom solvers for this eval in [./solvers.py](./solvers.py)
+
+1. `ExplicitStateSolver`: A nested solver that injects an explicit
+   representation of the task state after each number is seen. For example, for
+   the median task we inject the sorted list of numbers seen so far. For the
+   mode task, we inject a dictionary that maps each number seen so far to its
+   count. We view this solver as a baseline for the task, providing the
+   performance of the models on _explicit_ state tracking, rather than the
+   default _implicit_ state tracking.
+2. `RandomBaselineSolver`: A solver that randomly chooses a number from the
+   numbers seen so far as the rolling median or mode. In case of even length
+   lists in the median variant, it chooses two random numbers and returns their
+   arithmetic mean. We view this baseline as equivalent to randomly guessing.
+3. `TrackTheStatHuman`: A helper solver class that wraps the `HumanCliSolver`
+   class such that users do not have to wrap their answer in the
+   `[median: <answer>]` or `[mode: <answer>]` format and can instead just
+   directly type the number.
+
+## Token Usage Estimates
+
+Below are token usage estimates for a given run (one run = all samples) of the
+eval.
+
+For the mode task:
+
+| Model (state tracking)        | Input     | Output    | Total      |
+| ----------------------------- | --------- | --------- | ---------- |
+| gpt-3.5-turbo-0125 (implicit) | 670,000   | 10,000    | 680,000    |
+| gpt-3.5-turbo-0125 (explicit) | 2,710,000 | 30,000    | 2,740,000  |
+| gpt-4-base (implicit)         | 9,030,000 | 2,110,000 | 11,150,000 |
+| gpt-4-base (explicit)         | 3,720,000 | 960,000   | 4,680,000  |
+| gpt-4-0125-preview (implicit) | 3,050,000 | 30,000    | 3,080,000  |
+| gpt-4-0125-preview (explicit) | 8,580,000 | 50,000    | 8,630,000  |
+
+For the median task:
+
+| Model (state tracking)        | Input     | Output  | Total     |
+| ----------------------------- | --------- | ------- | --------- |
+| gpt-3.5-turbo-0125 (implicit) | 430,000   | 10,000  | 440,000   |
+| gpt-3.5-turbo-0125 (explicit) | 880,000   | 10,000  | 890,000   |
+| gpt-4-base (implicit)         | 2,900,000 | 760,000 | 3,660,000 |
+| gpt-4-base (explicit)         | 3,250,000 | 810,000 | 4,060,000 |
+| gpt-4-0125-preview (implicit) | 690,000   | 10,000  | 700,000   |
+| gpt-4-0125-preview (explicit) | 1,430,000 | 20,000  | 1,450,000 |
+
+## Future modifications
+
+- Identify new variants of the task beyond median or mode, where the explicit
+  state is either impossible to represent or not useful for the task. This would
+  allow us to more comfortably measure the implicit state tracking, even on CoT
+  solvers.
+- Identify more realistic and/or complex tasks.
+- Introduce distractors.
+
+## Version History
+
+- v0: Initial version released
+
+## Contribution Statement
+
+Eval design, implementation, and results evaluation were primarily conducted by
+Giulio Starace, under the guidance of (alphabetically by last-name) Steven
+Adler, Andrei Alexandru, James Aung, and Chan Jun Shern who provided research
+input, report revisions, and project management support.

evals/elsuite/track_the_stat/eval.py

@@ -0,0 +1,96 @@
+import logging
+import random
+from typing import Any, Optional
+
+import numpy as np
+
+from evals.elsuite.track_the_stat import prompts, utils
+from evals.eval import SolverEval
+from evals.record import RecorderBase, record_metrics
+from evals.solvers.solver import Solver
+from evals.task_state import Message, TaskState
+
+logging.getLogger("httpx").setLevel(logging.WARNING)
+logger = logging.getLogger(__name__)
+
+
+class TrackTheStat(SolverEval):
+    def __init__(self, task: str, n_samples: Optional[int] = 250, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert task in [
+            "median",
+            "mode",
+        ], f"task must be either 'median' or 'mode', but got {task}"
+        self.task = task
+        # warn, color in yellow
+        logger.warning(
+            utils.yellow_string(
+                "By nature of what is being evaluated, this eval assumes that the "
+                "solver cannot make use of external scratchpads or similar solutions "
+                "to explicitly write down the task state at every step. Using solvers "
+                "that allow for this functionality will likely produce invalid results."
+            )
+        )
+        self.task_desc = prompts.TASK_DESCRIPTION.format(
+            task=task,
+            task_further_details=prompts.task_to_further_details[task],
+            task_example=prompts.task_to_example[task],
+        )
+        self.task_fn = utils.task_to_fn[task]
+        self.n_samples = n_samples
+        self.rng = random.Random(self.seed)
+
+    def eval_sample(self, solver: Solver, sample: Any, rng: random.Random) -> None:
+        capped_inf_list = np.random.default_rng(sample["seed"]).integers(0, 100, size=300)
+        metrics = self._eval_sample(solver, capped_inf_list)
+
+        record_metrics(**metrics)
+
+    def _eval_sample(self, solver: Solver, capped_inf_list: list[int]) -> dict:
+        violation = False
+        task_state = TaskState(task_description=self.task_desc, messages=[])
+        for i, num in enumerate(capped_inf_list):
+            curr_list = capped_inf_list[: i + 1]
+            task_state.messages.append(Message(role="user", content=str(num)))
+            task_state.current_state = utils.compute_state(curr_list, self.task)
+            solver_output = solver(task_state).output
+            solver_response = utils.parse_solver_output(solver_output, self.task)
+            if solver_response is None:
+                violation = True
+                break
+            if round(solver_response, 1) != round(self.task_fn(curr_list), 1):
+                break
+            task_state.messages.append(Message(role="assistant", content=solver_output))
+
+        return {
+            "max_length": len(curr_list) - 1,
+            "violation": violation,
+        }
+
+    def run(self, recorder: RecorderBase):
+        samples = self._get_samples()
+        self.eval_all_samples(recorder, samples)
+        logged_metrics: list[dict] = recorder.get_metrics()
+
+        agg_metrics = self._compute_agg_metrics(logged_metrics)
+        return agg_metrics
+
+    def _compute_agg_metrics(self, logged_metrics: list[dict]) -> dict:
+        max_lengths = np.array([x["max_length"] for x in logged_metrics])
+
+        agg_metrics = {
+            "avg_max_length": np.mean(max_lengths),
+            "stddev_max_length": np.std(max_lengths),
+            "median_max_length": np.median(max_lengths),
+            "max_max_length": np.max(max_lengths),
+            "min_max_length": np.min(max_lengths),
+            "violation_rate": np.mean([x["violation"] for x in logged_metrics]),
+        }
+        # necessary for serialization, json doesn't like np floats
+        agg_metrics = {k: float(v) for k, v in agg_metrics.items()}
+        return agg_metrics
+
+    def _get_samples(self) -> list[dict]:
+        samples = [{"seed": num} for num in range(self.n_samples)]
+        samples = self.rng.sample(samples, self.n_samples)  # shuffling
+        return samples

evals/elsuite/track_the_stat/prompts/__init__.py

@@ -0,0 +1,27 @@
+from . import mode
+from . import median
+
+
+TASK_DESCRIPTION = """\
+In the following messages, you will be given a sequence of numbers.\
+ At each turn, you will be shown a number as input, and you should respond with the\
+ {task} of all the input numbers shown to you so far.
+
+{task_further_details}
+
+Here is an example of what this may look like.
+{task_example}
+
+Format your response as [{task}: <response>] (square brackets included), as shown in\
+the transcript above. The task will begin now.
+"""
+
+task_to_example = {
+    "median": median.MEDIAN_EXAMPLE,
+    "mode": mode.MODE_EXAMPLE,
+}
+
+task_to_further_details = {
+    "median": median.MEDIAN_FURTHER_DETAILS,
+    "mode": mode.MODE_FURTHER_DETAILS,
+}

evals/elsuite/track_the_stat/prompts/median.py

@@ -0,0 +1,33 @@
+MEDIAN_EXAMPLE = """\
+```example
+input: 1
+ideal_response: [median: 1]\
+ # your response; 1 is the only number shown so far
+---
+input: 2
+ideal_response: [median: 1.5]\
+ # even number of numbers, so median = mean(1,2) = 1.5
+---
+input: 1
+ideal_response: [median: 1]\
+ # 1 is now the middle number when sorting the numbers
+---
+input: 3
+ideal_response: [median: 1.5]\
+ # middle numbers are now 1 and 2, so once again median = mean(1,2) = 1.5
+---
+input: 3
+ideal_response: [median: 2]\
+# the sorted list is [1 1 2 3 3]; odd length, so median is the middle number, 2
+---
+input: 0
+ideal_response: [median: 1.5]\
+# the sorted list is [0 1 1 2 3 3]; even length, so median is mean(1,2) = 1.5
+```\
+"""
+
+
+MEDIAN_FURTHER_DETAILS = """\
+NOTE: In case of lists containing an even number of elements, you should respond with the\
+ arithmetic mean of the middle two numbers of the sorted list.\
+"""

evals/elsuite/track_the_stat/prompts/mode.py

@@ -0,0 +1,29 @@
+MODE_EXAMPLE = """\
+```example
+input: 1
+ideal_response: [mode: 1]\
+ # your response; 1 is the only number shown so far
+---
+input: 2
+ideal_response: [mode: 2]\
+ # 1 and 2 are tied modes (both appeared once), 2 > 1
+---
+input: 1
+ideal_response: [mode: 1]\
+ # 1 now has appeared more than any other number
+---
+input: 3
+ideal_response: [mode: 1]
+---
+input: 3
+ideal_response: [mode: 3]\
+ # 3 is tied with 1 in terms of appearances, 3 > 1
+---
+input: 0
+ideal_response: [mode: 3]
+```\
+"""
+
+MODE_FURTHER_DETAILS = """\
+NOTE: In case of ties, you should respond with the largest number that is part of the tie.\
+"""