src/utils.py

import pandas as pd
from pathlib import Path
from datasets import load_dataset
import numpy as np
import os
import re

UNVERIFIED_MODELS = [
    "nvidia/Nemotron-4-340B-Reward",
    "nvidia/Llama3-70B-SteerLM-RM",
    "Cohere May 2024",
    "google/gemini-1.5-pro-0514",
    "google/flame-24b-july-2024",
    "Cohere March 2024",
    "facebook/Self-taught-Llama-3-70B",
    "facebook/Self-taught-evaluator-llama3.1-70B",
    "google/flame-1.0-24B-july-2024",
    "Salesforce/SFR-LLaMa-3.1-70B-Judge-r",
    "Salesforce/SFR-nemo-12B-Judge-r",
    "Salesforce/SFR-LLaMa-3.1-8B-Judge-r",
    "SF-Foundation/TextEval-OffsetBias-12B",
    "SF-Foundation/TextEval-Llama3.1-70B",
    "nvidia/Llama-3.1-Nemotron-70B-Reward",
]

CONTAMINATED_MODELS = [
    "Skywork/Skywork-Reward-Gemma-2-27B",
    "Skywork/Skywork-Critic-Llama-3.1-70B",
    "LxzGordon/URM-LLaMa-3.1-8B",
    "Skywork/Skywork-Reward-Llama-3.1-8B",
    "Ray2333/GRM-Llama3-8B-rewardmodel-ft",
    "nicolinho/QRM-Llama3.1-8B",
    "nicolinho/QRM-Llama3-8B",
    "general-preference/GPM-Llama-3.1-8B",
    "general-preference/GPM-Gemma-2B"
]

# From Open LLM Leaderboard
def model_hyperlink(link, model_name):
    # if model_name is above 50 characters, return first 47 characters and "..."
    if len(model_name) > 50:
        model_name = model_name[:47] + "..."
    if model_name == "random":
        output = "random"
    elif model_name == "Cohere March 2024":
        output = f'<a target="_blank" href="https://huggingface.co/Cohere" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'
    elif "openai" == model_name.split("/")[0]:
        output = f'<a target="_blank" href="https://huggingface.co/openai" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'
    elif "Anthropic" == model_name.split("/")[0]:
        output = f'<a target="_blank" href="https://huggingface.co/Anthropic" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'
    elif "google" == model_name.split("/")[0]:
        output = f'<a target="_blank" href="https://huggingface.co/google" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'
    elif "PoLL" == model_name.split("/")[0]:
        output = model_name
    output = f'<a target="_blank" href="{link}" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'

    if model_name in UNVERIFIED_MODELS:
        output += " *"
    if model_name in CONTAMINATED_MODELS:
        output += " ⚠️"
    return output

def undo_hyperlink(html_string):
    # Regex pattern to match content inside > and <
    pattern = r'>[^<]+<'
    match = re.search(pattern, html_string)
    if match:
        # Extract the matched text and remove leading '>' and trailing '<'
        return match.group(0)[1:-1]
    else:
        return "No text found"


# Define a function to fetch and process data
def load_all_data(data_repo, subdir:str, subsubsets=False):    # use HF api to pull the git repo
    dir = Path(data_repo)
    data_dir = dir / subdir
    orgs = [d for d in os.listdir(data_dir) if os.path.isdir(os.path.join(data_dir, d))]
    # get all files within the sub folders orgs
    models_results = []
    for org in orgs:
        org_dir = data_dir / org
        files = [f for f in os.listdir(org_dir) if os.path.isfile(os.path.join(org_dir, f))]
        for file in files:
            if file.endswith(".json"):
                models_results.append(org + "/" + file)

    # create empty dataframe to add all data to
    df = pd.DataFrame()

    # load all json data in the list models_results one by one to avoid not having the same entries
    for model in models_results:
        model_data = load_dataset("json", data_files=data_repo + subdir+ "/" + model, split="train")
        df2 = pd.DataFrame(model_data)
        # add to df
        df = pd.concat([df2, df])


    # remove chat_template comlumn
    df = df.drop(columns=["chat_template"])

    # sort columns alphabetically
    df = df.reindex(sorted(df.columns), axis=1)
    
    # move column "model" to the front
    cols = list(df.columns)
    cols.insert(0, cols.pop(cols.index('model')))
    df = df.loc[:, cols]

    # select all columns except "model"
    cols = df.columns.tolist()
    cols.remove("model")
    # if model_type is a column (pref tests may not have it)
    if "model_type" in cols:
        cols.remove("model_type")
    # remove ref_model if in columns
    if "ref_model" in cols:
        cols.remove("ref_model")
    # remove model_beaker from dataframe
    if "model_beaker" in cols:
        cols.remove("model_beaker")
        df = df.drop(columns=["model_beaker"])
    
    # remove column xstest (outdated data)
    # if xstest is a column
    if "xstest" in cols:
        df = df.drop(columns=["xstest"])
        cols.remove("xstest")

    if "ref_model" in df.columns:
        df = df.drop(columns=["ref_model"])

    # remove column anthropic and summarize_prompted (outdated data)
    if "anthropic" in cols:
        df = df.drop(columns=["anthropic"])
        cols.remove("anthropic")
    if "summarize_prompted" in cols:
        df = df.drop(columns=["summarize_prompted"])
        cols.remove("summarize_prompted")
    # remove pku_better and pku_safer (removed from the leaderboard)
    if "pku_better" in cols:
        df = df.drop(columns=["pku_better"])
        cols.remove("pku_better")
    if "pku_safer" in cols:
        df = df.drop(columns=["pku_safer"])
        cols.remove("pku_safer")

    # convert to score 
    df[cols] = (df[cols]*100)
    avg = np.nanmean(df[cols].values,axis=1)
    # add average column
    df["average"] = avg
    
    # apply model_hyperlink function to column "model"
    df["model"] = df["model"].apply(lambda x: model_hyperlink(f"https://huggingface.co/{x}", x))

    # move average column to the second
    cols = list(df.columns)
    cols.insert(1, cols.pop(cols.index('average')))
    df = df.loc[:, cols]

    # move model_type column to first
    if "model_type" in cols:
        cols = list(df.columns)
        cols.insert(1, cols.pop(cols.index('model_type')))
        df = df.loc[:, cols]

    # remove models with DPO Ref. Free as type (future work)
    df = df[~df["model_type"].str.contains("DPO Ref. Free", na=False)]

    return df