from dataclasses import dataclass from enum import Enum @dataclass class Task: benchmark: str metric: str col_name: str # Select your tasks here # --------------------------------------------------- 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") class TasksMultimodal(Enum): task0 = Task("blimp", "acc", "BLiMP") task1 = Task("blimp_supplement", "acc", "BLiMP Supplement") task2 = Task("glue", "acc", "(Super)GLUE") task3 = Task("ewok", "acc", "EWoK") task4 = Task("vqa", "acc", "VQA") task5 = Task("winoground", "acc", "Winoground") task6 = Task("devbench", "acc", "DevBench") @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 # MCQA # models: list[str] # List of all models # layers: list[str] # 0-31 # col_name: str # display name in leaderboard # interventions: list[str] # output_token, output_location # counterfactuals: list[str] # symbol_counterfactual, etc. # metrics: list[str] # score # class TasksMib_Causalgraph(Enum): # task0 = TaskMIB_Causalgraph( # "MCQA", # ["LlamaForCausalLM", "Qwen2ForCausalLM", "Gemma2ForCausalLM"], # Updated model list # [str(i) for i in range(32)], # 0-31 layers # "mcqa", # ["output_token", "output_location"], # ["symbol_counterfactual", "randomLetter_counterfactual", # "answerPosition_counterfactual", "answerPosition_symbol_counterfactual"], # ["score"] # ) # class TasksMib_Causalgraph(Enum): # task0 = TaskMIB_Causalgraph( # "MCQA", # ["Qwen2ForCausalLM", "Gemma2ForCausalLM", "LlamaForCausalLM"], # Match exact model names with correct casing # [str(i) for i in range(32)], # "mcqa", # ["output_token", "output_location"], # ["randomLetter_counterfactual", "answerPosition_counterfactual", # "answerPosition_randomLetter_counterfactual"], # ["score"] # ) # class TasksMib_Causalgraph(Enum): # task0 = TaskMIB_Causalgraph( # "MCQA", # ["qwen2forcausallm", "gemma2forcausallm", "llamaforcausallm"], # Use lowercase consistently # [str(i) for i in range(32)], # "mcqa", # ["output_token", "output_location"], # ["randomLetter_counterfactual", "answerPosition_counterfactual", # "answerPosition_randomLetter_counterfactual"], # ["score"] # ) @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"] # ) class TasksMib_Causalgraph(Enum): task0 = TaskMIB_Causalgraph( "MCQA", ["qwen2forcausallm", "gemma2forcausallm", "llamaforcausallm"], # Use lowercase names to match actual columns { "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"] ) NUM_FEWSHOT = 0 # Change with your few shot # --------------------------------------------------- # Your leaderboard name TITLE = """

Mechanistic Interpretability Benchmark 2024 Leaderboards

""" # What does your leaderboard evaluate? INTRODUCTION_TEXT = """ The leaderboards for each track of the 2024 Mechanistic Interpretability Benchmark. """ # Which evaluations are you running? how can people reproduce what you have? LLM_BENCHMARKS_TEXT = f""" This leaderboard displays scores from the 2024 BabyLM Challenge. Each track has its own tab. """ EVALUATION_QUEUE_TEXT = """ ## Some good practices before requesting a predictions upload: Make sure you can get scores from your predictions file using the `score_predictions.py` script. ```bash git clone https://github.com/babylm/evaluation-pipeline-2024/ cd evaluation-pipeline-2024 python score_predictions.py path/to/your/predictions.json.gz ``` If this step fails, follow the error messages to debug your predictions before getting in touch. It's likely that either (i) some results are missing, or (ii) the results are incorrectly formatted. Make sure your model has an open license! This is a leaderboard that is meant to advance research on language modeling, and we'd love for as many people as possible to know they can use your model. Once these steps have been followed, get in touch with the organizers with your predictions file(s), and the scores you've obtained. We'll verify that we can match your scores, and then upload to the leaderboard. Optionally, you can give us your preferred model display name for the leaderboard, and a link to your model on HuggingFace. """ CITATION_BUTTON_LABEL = "If you would like to cite these results, please cite the 2024 BabyLM Findings paper, as well as the authors of the model(s) whose results you cite!" CITATION_BUTTON_TEXT = r""" @article{hu2024findingssecondbabylmchallenge, title={Findings of the Second BabyLM Challenge: Sample-Efficient Pretraining on Developmentally Plausible Corpora}, author={Michael Y. Hu and Aaron Mueller and Candace Ross and Adina Williams and Tal Linzen and Chengxu Zhuang and Ryan Cotterell and Leshem Choshen and Alex Warstadt and Ethan Gotlieb Wilcox}, year={2024}, journal={Computing Research Repository}, volume={arXiv:2412.05149}, url={https://arxiv.org/abs/2412.05149}, } """