Merge pull request rllm-org#146 from farrelmahaztra/musr

MuSR Benchmark Implementation

Author: alexandraabbas

README.md

@@ -349,6 +349,13 @@ The questions were generated by GPT-4 based on the "Computer Systems Security: P
    inspect eval inspect_evals/ifeval
    ```
 
+- ### [MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning](src/inspect_evals/musr)
+  Evaluating models on multistep soft reasoning tasks in the form of free text narratives.
+ <sub><sup>Contributed by: [@farrelmahaztra](https://github.com/farrelmahaztra)</sub></sup>
+   ```
+   inspect eval inspect_evals/musr
+   ```
+
 - ### [PAWS: Paraphrase Adversaries from Word Scrambling](src/inspect_evals/paws)
   Evaluating models on the task of paraphrase detection by providing pairs of sentences that are either paraphrases or not.
  <sub><sup>Contributed by: [@meltemkenis](https://github.com/meltemkenis)</sub></sup>

src/inspect_evals/_registry.py

@@ -53,6 +53,7 @@
 from .mmlu import mmlu_0_shot, mmlu_5_shot
 from .mmlu_pro import mmlu_pro
 from .mmmu import mmmu_multiple_choice, mmmu_open
+from .musr import musr
 from .paws import paws
 from .piqa import piqa
 from .pubmedqa import pubmedqa

src/inspect_evals/musr/README.md

@@ -0,0 +1,70 @@
+# MuSR
+
+[MuSR](https://arxiv.org/abs/2310.16049) is a dataset designed for evaluating models on multistep soft reasoning tasks in the form of free text narratives. It is composed of 3 domains: murder_mysteries, object_placements, and team_allocation, which have 250, 256, and 250 instances respectively in the dataset.
+
+<!-- Contributors: Automatically Generated -->
+Contributed by [@farrelmahaztra](https://github.com/farrelmahaztra)
+<!-- /Contributors: Automatically Generated -->
+
+<!-- Usage: Automatically Generated -->
+## Usage
+
+First, install the `inspect_ai` and `inspect_evals` Python packages with:
+```bash
+pip install inspect_ai
+pip install git+https://github.com/UKGovernmentBEIS/inspect_evals
+```
+
+Then, evaluate against one or more models with:
+```bash
+inspect eval inspect_evals/musr --model openai/gpt-4o
+```
+
+After running evaluations, you can view their logs using the `inspect view` command:
+
+```bash
+inspect view
+```
+
+If you don't want to specify the `--model` each time you run an evaluation, create a `.env` configuration file in your working directory that defines the `INSPECT_EVAL_MODEL` environment variable along with your API key. For example:
+
+```bash
+INSPECT_EVAL_MODEL=anthropic/claude-3-5-sonnet-20240620
+ANTHROPIC_API_KEY=<anthropic-api-key>
+```
+<!-- /Usage: Automatically Generated -->
+
+<!-- Options: Automatically Generated -->
+## Options
+
+You can control a variety of options from the command line. For example:
+```bash
+inspect eval inspect_evals/musr --limit 10
+inspect eval inspect_evals/musr --max-connections 10
+inspect eval inspect_evals/musr --temperature 0.5
+```
+
+See `inspect eval --help` for all available options.
+<!-- /Options: Automatically Generated -->
+
+## Dataset
+Here is a truncated example from the dataset:
+
+>The tension between them was palpable. Alice had been awarded a major journalist award that Gabrielle had desired. This only deepened their rivalry, with Gabrielle feeling overlooked for this recognition in the Jazz scene.
+>
+>Winston cast his gaze over the club once more—a hub of pulsating rhythms now eerily silent.
+>
+>A significant part of the evening was Gabrielle's recorded interview with Alice. It played on the local radio, their professional rivalry subtly echoing under their professional demeanor.
+>
+>With a deep breath, Winston knew he had a tall task ahead. The jazz club, where Alice was last seen alive was now shrouded in an eerie silence, the vibrant rhythms of what used to be a lively night echoing in the abandoned stage. It was up to him to piece together the missing notes and bring the symphony of this unsolved case to a satisfying finale.
+>
+>Who is the most likely murderer?
+>
+>Pick one of the following choices:
+>A - Eugene
+>B - Gabrielle 
+
+The model is tasked to answer the question and choose the appropriate option.
+
+## Evaluation
+The prompts are based on the [official MuSR repository](https://github.com/Zayne-sprague/MuSR). The in-built `multiple_choice` scorer is used for evaluation.

src/inspect_evals/musr/__init__.py

@@ -0,0 +1,3 @@
+from .musr import musr
+
+__all__ = ["musr"]

src/inspect_evals/musr/musr.py

@@ -0,0 +1,124 @@
+"""MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning
+
+Zayne Sprague, Xi Ye, Kaj Bostrom, Swarat Chaudhuri, and Greg Durrett.
+https://arxiv.org/abs/2310.16049
+
+# Example: Eval MuSR with team_allocation and CoT+
+inspect eval musr.py -T domain=team_allocation -T prompt_technique=cot+
+
+"""
+
+import ast
+from typing import Any, Dict, Literal
+
+from inspect_ai import Task, task
+from inspect_ai.dataset import Sample, hf_dataset
+from inspect_ai.scorer import choice
+from inspect_ai.solver import multiple_choice, system_message
+
+from inspect_evals.musr.prompts import (
+    COT_PLUS_PROMPT,
+    COT_PROMPT,
+    MURDER_MYSTERIES_EXAMPLE,
+    MURDER_MYSTERIES_HINT,
+    OBJECT_PLACEMENTS_EXAMPLE,
+    OBJECT_PLACEMENTS_HINT,
+    REGULAR_PROMPT,
+    SYSTEM_PROMPT,
+    TEAM_ALLOCATION_EXAMPLE,
+    TEAM_ALLOCATION_HINT,
+)
+
+DomainType = Literal["murder_mysteries", "object_placements", "team_allocation"]
+PromptTechniqueType = Literal["regular", "cot", "cot+"]
+
+DEFAULT_DOMAIN: DomainType = "murder_mysteries"
+DEFAULT_PROMPT_TECHNIQUE: PromptTechniqueType = "regular"
+DEFAULT_EXAMPLE_COUNT = 0
+
+
+@task
+def musr(
+    domain: DomainType = DEFAULT_DOMAIN,
+    prompt_technique: PromptTechniqueType = DEFAULT_PROMPT_TECHNIQUE,
+    example_count: int = DEFAULT_EXAMPLE_COUNT,
+) -> Task:
+    """Inspect task implementing the MuSR benchmark.
+
+    Args:
+        domain (Literal["murder_mysteries", "object_placements", "team_allocation"]): Which domain in the dataset to evaluate.
+            Defaults to "murder_mysteries".
+        prompt_technique (Literal["regular", "cot", "cot+"]): The prompt technique to use. "regular" includes only the narrative
+            and the question. "cot" uses chain-of-thought prompting. "cot+" includes a hint. Defaults to "regular".
+        example_count (int): Number of solved examples to include at the beginning of each prompt. Defaults to 0. Currently only supports 1 example.
+    """
+    prompt = get_domain_prompt(domain, prompt_technique, example_count)
+
+    dataset = hf_dataset(
+        path="TAUR-Lab/MuSR",
+        split=domain,
+        sample_fields=record_to_sample,
+        shuffle=True,
+        auto_id=True,
+    )
+
+    return Task(
+        dataset=dataset,
+        solver=[
+            system_message(SYSTEM_PROMPT),
+            multiple_choice(template=prompt),
+        ],
+        scorer=choice(),
+    )
+
+
+def get_domain_prompt(
+    domain: DomainType = DEFAULT_DOMAIN,
+    prompt_technique: PromptTechniqueType = DEFAULT_PROMPT_TECHNIQUE,
+    example_count: int = DEFAULT_EXAMPLE_COUNT,
+) -> str:
+    domain_info = {
+        "murder_mysteries": {
+            "hint": MURDER_MYSTERIES_HINT,
+            "example": MURDER_MYSTERIES_EXAMPLE,
+        },
+        "object_placements": {
+            "hint": OBJECT_PLACEMENTS_HINT,
+            "example": OBJECT_PLACEMENTS_EXAMPLE,
+        },
+        "team_allocation": {
+            "hint": TEAM_ALLOCATION_HINT,
+            "example": TEAM_ALLOCATION_EXAMPLE,
+        },
+    }
+
+    if domain not in domain_info:
+        raise ValueError(
+            "Unknown domain. Valid domains are murder_mysteries (default), object_placements, and team_allocation"
+        )
+
+    if prompt_technique == "regular":
+        prompt = REGULAR_PROMPT
+    elif prompt_technique == "cot":
+        prompt = COT_PROMPT
+    elif prompt_technique == "cot+":
+        prompt = COT_PLUS_PROMPT.format(hint=domain_info[domain]["hint"])
+    else:
+        raise ValueError(
+            "Unknown prompt technique. Valid prompt techniques are regular (default), cot, and cot+."
+        )
+
+    if example_count > 1:
+        raise ValueError(">1 examples currently not supported")
+    if example_count == 1:
+        return f'Here is an example of solving the task:\n\n{domain_info[domain]["example"]}\n\nThis is the end of the example. The real task is below.\n\n---\n\n{prompt}'
+    else:
+        return prompt
+
+
+def record_to_sample(record: Dict[str, Any]) -> Sample:
+    return Sample(
+        input=f'{record["narrative"]}\n\n{record["question"]}',
+        choices=ast.literal_eval(record["choices"]),
+        target=chr(ord("A") + int(record["answer_index"])),
+    )