Error Recovery Eval (openai#1485)

@JunShern is going to 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

Error Recovery

### Eval description

Evaluates how well models can recovery from mistakes

### 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>

---------

Co-authored-by: Ian McKenzie <[email protected]>

evals/elsuite/error_recovery/README.md

@@ -0,0 +1,96 @@
+# Error Recovery
+
+This eval evaluates how well LMs can recover from having made reasoning errors while solving a problem.
+
+## Usage
+
+Run with:
+
+```
+oaieval <solver> error-recovery
+```
+
+We suggest using `error_recovery/gpt-4-0613` or `error_recovery/gpt-3.5-turbo-0613` as default choices for `<solver>`
+
+For more examples of running this eval, see `scripts/run_experiments.sh`
+
+## Dataset
+
+The dataset consists of 1436 problems from [BIG-Bench Mistake](https://github.com/WHGTyen/BIG-Bench-Mistake) which have reasoning traces with mistakes in them, and incorrect answers.
+
+## Evaluation Process
+
+For each example, we run the solver on three variations of the problem:
+
+1) No Reasoning (NR) variation - Give the solver the question and ask it to generate reasoning and an answer. Example below:
+```
+[SYSTEM]
+Solve the given problem, writing your reasoning along the way.
+
+[USER]
+Solve this problem for x: x = ((3 + 3) * 5) + 2
+```
+
+
+2) Correct Reasoning (CR) variation - Give the solver the question and some correct reasoning steps and have it keep generating reasoning and an answer. Example below:
+
+```
+[SYSTEM]
+Solve the given problem, writing your reasoning along the way.
+
+[USER]
+Solve this problem for x: x = ((3 + 3) * 5) + 2
+
+[ASSISTANT]
+3 + 3 = 6
+```
+
+3) Incorrect Reasoning (IR) variation - Give the solver the question, the same correct reasoning steps as before, and an incorrect reasoning step and have it keep generating reasoning and an answer. 
+
+```
+[SYSTEM]
+Solve the given problem, writing your reasoning along the way.
+
+[USER]
+Solve this problem for x: x = ((3 + 3) * 5) + 2
+
+[ASSISTANT]
+3 + 3 = 6
+6 * 5 = 35
+```
+
+
+## Metrics
+
+The below are the key metrics of this eval:
+
+| Metric | Interpretation |
+| --- | --- |
+| `NR_correct_rate` | The accuracy of the solver in the NR variation.
+| `CR_correct_rate` | The accuracy of the solver in the CR variation.
+| `IR_correct_rate` | The accuracy of the solver in the IR variation.
+
+## Variants
+
+| Variant | Notes |
+| --- | --- |
+| Default: `error-recovery.main` | Default setting. Includes all 1436 examples. |
+| `error-recovery.medium` | A balanced subset of 250 examples, 50 from each task. |
+| `error-recovery.small` | A balanced subset of 10 examples, 2 from each task (primarily for quick tests). |
+
+## Token Usage Estimates
+
+Below is a rough estimate of the total number of tokens consumed by the default variant:
+
+| Solver | Tokens |
+| --- | --- |
+| error_recovery/gpt-3.5-turbo-0613 | 4,750,781 |
+| error_recovery/gpt-4-0613 | 4,386,440 |
+
+## Version History
+
+- v0: Initial version released
+
+## Contribution statement
+
+Eval design, implementation, results evaluation, and report were primarily done by Ian McKenzie based on an initial version by Oliver Jaffe. Work done under the guidance of (alphabetically by last-name) Steven Adler, James Aung, and Chan Jun Shern, who provided research input, report revisions, and project management support.

evals/elsuite/error_recovery/defaults.py

@@ -0,0 +1,15 @@
+DEFAULT_TASK_DESCRIPTION = "Solve the given problem, writing your reasoning along the way."
+
+DEFAULT_MISTAKE_MESSAGE = "There might be a mistake in your reasoning."
+
+DEFAULT_FINAL_ANSWER_MESSAGE = (
+ "Given this reasoning, write your final answer. Only write your final answer, and nothing else."
+)
+
+TASK_SPECIFIC_EXTRACTION_INFO = {
+ "dyck_languages": "\n\nAnswer with just the end of the sequence, separated by spaces. Do not repeat the part of the sequence given in the question. Only write the sequence of symbols, nothing else.",
+ "logical_deduction": "\n\nAnswer with the selected single letter indicating your answer, wrapped with parentheses. Do not write anything else.",
+ "multistep_arithmetic": "\n\nAnswer with a single number.",
+ "tracking_shuffled_objects": "\n\nAnswer with the selected single letter indicating your answer, wrapped with parentheses. Do not write anything else.",
+ "word_sorting": "\n\nAnswer with the sorted words, each lower case and separated by a single space.",
+}

evals/elsuite/error_recovery/eval.py

@@ -0,0 +1,284 @@
+import copy
+import random
+from dataclasses import dataclass
+from typing import Any, List, Literal, Optional, Sequence
+
+import evals
+import evals.metrics
+import evals.record
+from evals.api import CompletionFn
+from evals.elsuite.error_recovery.defaults import (
+ DEFAULT_FINAL_ANSWER_MESSAGE,
+ DEFAULT_MISTAKE_MESSAGE,
+ DEFAULT_TASK_DESCRIPTION,
+ TASK_SPECIFIC_EXTRACTION_INFO,
+)
+from evals.eval import SolverEval
+from evals.solvers.solver import Solver
+from evals.task_state import Message, TaskState
+
+# possible Mistake NOTIFiciation POSitions
+MistakeNotifPos = Literal["immediate", "end"]
+
+
+@dataclass
+class Sample:
+ question: str
+ correct_steps: Sequence[str]
+ incorrect_step: str
+ target: Any
+ task: str
+ num_ground_truth_steps: int
+ mistake_index: int
+
+
+class ErrorRecovery(SolverEval):
+ def __init__(
+ self,
+ completion_fns: Sequence[CompletionFn],
+ samples_jsonl: str,
+ n_samples: Optional[int] = None,
+ mistake_notification_position: Optional[MistakeNotifPos] = None,
+ mistake_notification_for_ir_only: bool = False,
+ mark_as_own_reasoning: bool = True,
+ final_answer_prompt_role: str = "system",
+ *args,
+ **kwargs,
+ ):
+ """Evaluate a solver on the error recovery task.
+
+ Args:
+ completion_fns: The completion functions to evaluate. (should be a single solver)
+ samples_jsonl: The relative path to the samples jsonl file in evals/registry/data.
+ n_samples: The number of samples to use. If None, use all samples.
+ mistake_notification_position: The position of the mistake
+ notification. Options are "immediate" for right after the provided
+ reasoning, or "end" for right after the model-generated reasoning.
+ If None, no mistake notification is added.
+ mistake_notification_for_ir_only: Whether to only add the mistake notification
+ for the incorrect reasoning case. If True, the mistake notification is
+ added for the incorrect reasoning case, and not for the correct reasoning
+ or no reasoning cases.
+ mark_as_own_reasoning: Whether to include the sample reasoning as an
+ 'assistant' or 'user' message.
+ final_answer_prompt_role: The role to use for the final answer prompt. Should
+ be either "system" or "user".
+ """
+ super().__init__(
+ completion_fns=completion_fns, samples_jsonl=samples_jsonl, *args, **kwargs
+ )
+
+ self.n_samples = n_samples
+ self.mistake_notif_pos: Optional[MistakeNotifPos] = mistake_notification_position
+ self.mistake_notif_ir_only = mistake_notification_for_ir_only
+
+ # there are some issues with passing bools in from extra_eval_params
+ assert isinstance(mark_as_own_reasoning, bool)
+ self.mark_as_own_reasoning = mark_as_own_reasoning
+
+ self.final_answer_prompt_role = final_answer_prompt_role
+ assert self.final_answer_prompt_role in ["system", "user"]
+
+ def eval_sample(self, solver: Solver, sample: Sample, rng: random.Random, extra_logging=None):
+ task = sample.task
+
+ # Get the baseline with no provided reasoning
+ nr_task_state = self._get_no_reasoning_task_state(sample)
+ # only "end" makes sense for 'no reasoning'
+ nr_notif_pos = "end" if self.mistake_notif_pos == "end" else None
+ if self.mistake_notif_ir_only:
+ nr_notif_pos = None
+
+ nr_answer = self._get_answer(
+ solver=solver,
+ task_state=nr_task_state,
+ sample=sample,
+ mistake_notif_pos=nr_notif_pos,
+ )
+
+ # Run with correct reasoning
+ cr_task_state = self._get_correct_reasoning_task_state(sample)
+ cr_notif_pos = self.mistake_notif_pos
+ if self.mistake_notif_ir_only:
+ cr_notif_pos = None
+
+ cr_answer = self._get_answer(
+ solver=solver,
+ task_state=cr_task_state,
+ sample=sample,
+ mistake_notif_pos=cr_notif_pos,
+ )
+
+ # Run with incorrect reasoning
+ ir_task_state = self._get_incorrect_reasoning_task_state(sample)
+ ir_notif_pos = self.mistake_notif_pos
+
+ ir_answer = self._get_answer(
+ solver=solver,
+ task_state=ir_task_state,
+ sample=sample,
+ mistake_notif_pos=ir_notif_pos,
+ )
+
+ assert len(sample.correct_steps) == sample.mistake_index
+
+ metrics = {
+ "task": task,
+ "num_ground_truth_steps": sample.num_ground_truth_steps,
+ "mistake_index": sample.mistake_index,
+ "target": str(sample.target), # ground truth answer
+ "mistake_notification_position": self.mistake_notif_pos,
+ "mistake_notification_for_ir_only": self.mistake_notif_ir_only,
+ "NR_sampled": nr_answer,
+ "CR_sampled": cr_answer,
+ "IR_sampled": ir_answer,
+ "NR_correct": nr_answer == str(sample.target),
+ "CR_correct": cr_answer == str(sample.target),
+ "IR_correct": ir_answer == str(sample.target),
+ }
+ evals.record.record_metrics(**metrics)
+
+ def _get_no_reasoning_task_state(self, sample: Sample) -> TaskState:
+ task_description = DEFAULT_TASK_DESCRIPTION
+ no_reasoning_messages = [
+ Message(role="user", content=sample.question),
+ ]
+ no_reasoning_task_state = TaskState(
+ task_description=task_description,
+ messages=no_reasoning_messages,
+ )
+ return no_reasoning_task_state
+
+ def _get_correct_reasoning_task_state(self, sample: Sample) -> TaskState:
+ task_description = DEFAULT_TASK_DESCRIPTION
+ correct_steps = "\n".join(sample.correct_steps)
+ reasoning_role = "assistant" if self.mark_as_own_reasoning else "user"
+ correct_reasoning_messages = [
+ Message(role="user", content=sample.question),
+ Message(role=reasoning_role, content=correct_steps),
+ ]
+ correct_reasoning_task_state = TaskState(
+ task_description=task_description,
+ messages=correct_reasoning_messages,
+ )
+ return correct_reasoning_task_state
+
+ def _get_incorrect_reasoning_task_state(
+ self,
+ sample: Sample,
+ ) -> TaskState:
+ task_description = DEFAULT_TASK_DESCRIPTION
+ correct_steps = "\n".join(sample.correct_steps)
+ steps_with_incorrect_reasoning = f"{correct_steps}\n{sample.incorrect_step}"
+ reasoning_role = "assistant" if self.mark_as_own_reasoning else "user"
+ incorrect_reasoning_messages = [
+ Message(role="user", content=sample.question),
+ Message(role=reasoning_role, content=steps_with_incorrect_reasoning),
+ ]
+
+ incorrect_reasoning_task_state = TaskState(
+ task_description=task_description,
+ messages=incorrect_reasoning_messages,
+ )
+ return incorrect_reasoning_task_state
+
+ def _get_answer(
+ self,
+ solver: Solver,
+ task_state: TaskState,
+ sample: Sample,
+ mistake_notif_pos: Optional[MistakeNotifPos],
+ ) -> str:
+ """Get a final answer from the solver for a given sample.
+
+ Args:
+ solver: The solver to use.
+ task_state: The task state to use.
+ sample: The Sample being evaluated (relevant for answer extraction).
+ mistake_notification_position: The position of the mistake notification.
+ Options are "immediate" for right after the provided reasoning, or "end" for right
+ after the model-generated reasoning. If None, no mistake notification is added.
+
+ TODO (ian): Work out whether to add mistake notification to 'no reasoning' baseline
+ """
+ mistake_message = Message("user", DEFAULT_MISTAKE_MESSAGE)
+ if mistake_notif_pos == "immediate":
+ task_state.messages.append(mistake_message)
+
+ output = solver(task_state=task_state).output
+ task_state.messages.append(Message("assistant", output))
+
+ # run solver again if mistake notification is at the end
+ if mistake_notif_pos == "end":
+ task_state.messages.append(mistake_message)
+ output = solver(task_state=task_state).output
+ task_state.messages.append(Message("assistant", output))
+
+ answer = self._extract_final_answer(solver=solver, task_state=task_state, sample=sample)
+ return answer
+
+ def run(self, recorder: evals.record.Recorder):
+ samples = self.get_samples()
+
+ self.eval_all_samples(recorder, samples)
+ metrics = recorder.get_metrics()
+
+ NR_correct_rate = len([i for i in metrics if i["NR_correct"]]) / len(metrics)
+ CR_correct_rate = len([i for i in metrics if i["CR_correct"]]) / len(metrics)
+ IR_correct_rate = len([i for i in metrics if i["IR_correct"]]) / len(metrics)
+
+ results = {
+ "NR_correct_rate": NR_correct_rate,
+ "CR_correct_rate": CR_correct_rate,
+ "IR_correct_rate": IR_correct_rate,
+ }
+
+ # Split results per type of task
+ all_tasks = set([i["task"] for i in metrics])
+ for task in all_tasks:
+ filtered_metrics = [i for i in metrics if i["task"] == task]
+ NR_correct_rate = len([i for i in filtered_metrics if i["NR_correct"]]) / len(
+ filtered_metrics
+ )
+ CR_correct_rate = len([i for i in filtered_metrics if i["CR_correct"]]) / len(
+ filtered_metrics
+ )
+ IR_correct_rate = len([i for i in filtered_metrics if i["IR_correct"]]) / len(
+ filtered_metrics
+ )
+
+ # we use hyphens in the task name so they can be extracted by splitting on underscores
+ task_string = task.replace("_", "-")
+ results.update(
+ {
+ f"task_{task_string}_NR_correct_rate": NR_correct_rate,
+ f"task_{task_string}_CR_correct_rate": CR_correct_rate,
+ f"task_{task_string}_IR_correct_rate": IR_correct_rate,
+ }
+ )
+
+ return results
+
+ def _extract_final_answer(self, solver: Solver, task_state: TaskState, sample: Sample):
+ """Extract the final answer from the solver output using the same solver."""
+ task_state = copy.deepcopy(task_state)
+
+ task_specific_info = TASK_SPECIFIC_EXTRACTION_INFO[sample.task]
+ final_answer_prompt = DEFAULT_FINAL_ANSWER_MESSAGE + task_specific_info
+
+ task_state.messages.append(
+ Message(role=self.final_answer_prompt_role, content=final_answer_prompt)
+ )
+ answer = solver(task_state=task_state).output
+
+ return answer
+
+ def get_samples(self) -> List[Sample]:
+ samples = super().get_samples()
+
+ if self.n_samples is not None:
+ assert (
+ len(samples) >= self.n_samples
+ ), f"Can't get {self.n_samples} samples from a dataset with {len(samples)} samples"
+ samples = samples[: self.n_samples]
+ return [Sample(**sample_dict) for sample_dict in samples]