src/about.py

from dataclasses import dataclass
from enum import Enum

@dataclass
class Task:
    benchmark: str
    metric: str
    col_name: str

class Tasks(Enum):
    # task_key in the json file, metric_key in the json file, name to display in the leaderboard 
    task0 = Task("blimp", "acc", "BLiMP")
    task1 = Task("blimp_supplement", "acc", "BLiMP Supplement")
    task2 = Task("glue", "acc", "(Super)GLUE")
    task3 = Task("ewok", "acc", "EWoK")

@dataclass
class TaskMIB_Subgraph:
    benchmark: str      # task name in json (ioi/arithmetic)
    models: list[str]   # list of models to show as sub-columns
    col_name: str       # display name in leaderboard
    metrics: list[str]  # metrics to store (edge_counts, faithfulness)

class TasksMib_Subgraph(Enum):
    task0 = TaskMIB_Subgraph("ioi", ["gpt2", "qwen2_5", "gemma2", "llama3"], "IOI", ["edge_counts", "faithfulness"])
    task1 = TaskMIB_Subgraph("mcqa", ["qwen2_5", "gemma2", "llama3"], "MCQA", ["edge_counts", "faithfulness"])
    task2 = TaskMIB_Subgraph("arithmetic_addition", ["llama3"], "arithmetic_addition", ["edge_counts", "faithfulness"])
    task3 = TaskMIB_Subgraph("arithmetic_subtraction", ["llama3"], "arithmetic_subtraction", ["edge_counts", "faithfulness"])
    task4 = TaskMIB_Subgraph("arc_easy", ["gemma2", "llama3"], "arc_easy", ["edge_counts", "faithfulness"])
    task5 = TaskMIB_Subgraph("arc_challenge", ["llama3"], "arc_challenge", ["edge_counts", "faithfulness"])

    @classmethod
    def get_all_tasks(cls):
        """Returns a list of all task benchmarks"""
        return [task.value.benchmark for task in cls]
    
    @classmethod
    def get_all_models(cls):
        """Returns a list of all unique models across all tasks"""
        models = set()
        for task in cls:
            models.update(task.value.models)
        return sorted(list(models))




# @dataclass 
# class TaskMIB_Causalgraph:
#     benchmark: str      
#     models: list[str]   
#     layers: dict[str, list[str]]  # Different layers for each model
#     col_name: str      
#     interventions: list[str]  
#     counterfactuals: list[str]  
#     metrics: list[str]  


# class TasksMib_Causalgraph(Enum):
#     task0 = TaskMIB_Causalgraph("MCQA", 
#         ["qwen2forcausallm", "gemma2forcausallm", "llamaforcausallm"],
#         {
#             "qwen2forcausallm": [str(i) for i in range(24)],    # 0-23
#             "gemma2forcausallm": [str(i) for i in range(26)],   # 0-25
#             "llamaforcausallm": [str(i) for i in range(32)]     # 0-31
#         },
#         "mcqa",
#         ["output_token", "output_location"],
#         ["randomLetter_counterfactual", "answerPosition_counterfactual", 
#          "answerPosition_randomLetter_counterfactual"],
#         ["score"]
#     )


@dataclass
class TaskMIB_Causalgraph:
    benchmark: str      # task name in json (ioi/arithmetic)
    models: list[str]   # list of models to show as sub-columns
    col_name: str       # display name in leaderboard
    metrics: list[str]  # metrics to store (average_score)
    target_variables: list[str]



class TasksMib_Causalgraph(Enum):
    task0 = TaskMIB_Causalgraph("ioi", ["Qwen2ForCausalLM", "GPT2ForCausalLM", "GPT2LMHeadModel", "Gemma2ForCausalLM", "LlamaForCausalLM"], "ioi_task", ["average_score"], ["output_token", "output_position"])
    task1 = TaskMIB_Causalgraph("mcqa", ["Qwen2ForCausalLM", "GPT2ForCausalLM", "GPT2LMHeadModel", "Gemma2ForCausalLM", "LlamaForCausalLM"], "4_answer_MCQA", ["average_score"], ["answer_pointer", "answer"])
    task2 = TaskMIB_Causalgraph("ravel", ["Qwen2ForCausalLM", "GPT2ForCausalLM", "GPT2LMHeadModel", "Gemma2ForCausalLM", "LlamaForCausalLM"], "RAVEL_task", ["average_score"], ["Continent", "Language", "Country", "Language"])
    task3 = TaskMIB_Causalgraph("arc_easy", ["Qwen2ForCausalLM", "GPT2ForCausalLM", "GPT2LMHeadModel", "Gemma2ForCausalLM", "LlamaForCausalLM"], "ARC_easy", ["average_score"], ["answer_pointer", "answer"])

    @classmethod
    def get_all_tasks(cls):
        """Returns a list of all task benchmarks"""
        return [task.value.benchmark for task in cls]
    
    @classmethod
    def get_all_models(cls):
        """Returns a list of all unique models across all tasks"""
        models = set()
        for task in cls:
            models.update(task.value.models)
        return sorted(list(models))



# Your leaderboard name
TITLE = """<h1 align="center" id="space-title"> Mechanistic Interpretability Benchmark Leaderboards</h1>"""

# What does your leaderboard evaluate?
INTRODUCTION_TEXT = """
The leaderboards for each track of the Mechanistic Interpretability Benchmark.
"""

# Which evaluations are you running? how can people reproduce what you have?
LLM_BENCHMARKS_TEXT = f"""
This leaderboard displays scores on the private test set for the Mechanistic Interpretability Benchmark. Each track has its own tab.
"""

EVALUATION_QUEUE_TEXT_SUBGRAPH = """
## Circuit localization track

### 1. Collect your circuits
You'll need either (i) 1 circuit per task/model combinaton with floating-point importance scores for each edge or node, or (ii) 9 circuits per model/task with binary membership scores for each edge or node.
For specifications about the file formats we accept, see the README on [our project GitHub](https://github.com/hannamw/MIB-subgraph-track).

### 2. Upload your circuits
Create a HuggingFace repository, and create a folder in that repository that will hold all of your circuit folders.
At the URL you provide, there should be one folder per task/model combination; these folders
should contain your circuit(s). As long as the folder names contain the model and task names, you do not need to worry about the circuit filenames.
If you provide more circuits than needed, our evaluation script will take the first 9 lexicographically in a given folder. We provide examples of valid
submissions: see [here](https://huggingface.co/mib-bench/mib-circuits-example/tree/main/importances/json) for a submission using importance scores and
[here](https://huggingface.co/mib-bench/mib-circuits-example/tree/main/multiple_circuits/pt) for a submission uploading multiple circuits.

### 3. Manage your submission in the queue
If your submission passes all checks, it will be added to the queue. You will receive a submission ID here when you do this; be sure to save it!
This will allow you to remove your submission from the queue (e.g., if you find a bug in your circuits). This will prevent you from needing to wait until
next week to resubmit.

Before your submission has been validated by our backend, it will have the "PREVALIDATION" status in the queue. Once it has been validated, it will have the "PENDING" status.
It will keep the PENDING status until it has been run on the private test set.
"""

EVALUATION_QUEUE_TEXT_CAUSALVARIABLE = """
## Causal variable localization track

### 1. Collect your materials
You'll need the following:
* Trained featurizer, inverse featurizer, and indices objects.
* A python file containing the implementation of your featurizer and inverse featurizer.
* (Optional) Dynamic token alignment functions, provided in another python file.

### 2. Upload your materials
Create a HuggingFace repository, and create a folder in that repository that will hold all of your materials.
At the URL you provide (we'll call this the "root"), each of the above materials should be present. At the linked folder,
we will take the first python script lexicographically at the root as the featurizer script. Within that folder, we expect
one subfolder per model/task/causal variable triplet. Each subfolder should contain the trained featurizer, inverse featurizer,
and indices.

### 3. Manage your submission in the queue
If your submission passes all checks, it will be added to the queue. You will receive a submission ID here when you do this; be sure to save it!
This will allow you to remove your submission from the queue (e.g., if you find a bug). This will prevent you from needing to wait until
next week to resubmit.

Before your submission has been validated by our backend, it will have the "PREVALIDATION" status in the queue. Once it has been validated, it will have the "PENDING" status.
It will keep the PENDING status until it has been run on the private test set.
"""

CITATION_BUTTON_LABEL = "If you would like to cite these results, please cite the MIB paper, as well as the author(s) of the method(s) whose results you cite!"
CITATION_BUTTON_TEXT = r"""@article{mib-2025,
	title = {{MIB}: A Mechanistic Interpretability Benchmark},
	author = {Aaron Mueller and Atticus Geiger and Sarah Wiegreffe and Dana Arad and Iv{\'a}n Arcuschin and Adam Belfki and Yik Siu Chan and Jaden Fiotto-Kaufman and Tal Haklay and Michael Hanna and Jing Huang and Rohan Gupta and Yaniv Nikankin and Hadas Orgad and Nikhil Prakash and Anja Reusch and Aruna Sankaranarayanan and Shun Shao and Alessandro Stolfo and Martin Tutek and Amir Zur and David Bau and Yonatan Belinkov},
	year = {2025},
	journal = {CoRR},
	volume = {arXiv:2504.13151},
	url = {https://arxiv.org/abs/2504.13151v1}
}
"""