from collections import defaultdict from dataclasses import dataclass from typing import Dict, List import numpy as np import pandas as pd from datasets import load_dataset from content import PLOT_1_TITLE, PLOT_2_TITLE, PLOT_3_TITLE, PLOT_4_TITLE from utils import make_clickable_model from visualizations import ( get_bootstrap_result, switch_model_a_b, visualize_battle_count, visualize_bootstrap_scores, visualize_pairwise_win_fraction, visualize_rating_count, ) KOALA_LINK = "https://huggingface.co/TheBloke/koala-13B-HF" VICUNA_LINK = "https://huggingface.co/HuggingFaceH4/stable-vicuna-13b-2904" OASST_LINK = "https://huggingface.co/OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5" DOLLY_LINK = "https://huggingface.co/databricks/dolly-v2-12b" MODEL_PAGE = "https://huggingface.co/models" def make_clickable_model_elo(model_name): link = "" if model_name == "dolly-12b": link = DOLLY_LINK elif model_name == "vicuna-13b": link = VICUNA_LINK elif model_name == "koala-13b": link = KOALA_LINK elif model_name == "oasst-12b": link = OASST_LINK else: link = MODEL_PAGE return f'{model_name}' @dataclass class EloEvalResult: model: str gpt_4_all: int human_all: int human_instruct: int human_code_instruct: int tie_allowed: bool def to_dict(self): base_model = f"{self.model}" data_dict = {} data_dict["Model"] = make_clickable_model_elo(base_model) data_dict["GPT-4 (all)"] = self.gpt_4_all data_dict["Human (all)"] = self.human_all data_dict["Human (instruct)"] = self.human_instruct data_dict["Human (code-instruct)"] = self.human_code_instruct return data_dict def create_eval_df(df, tie_allowed): responses = [] for _, row in df.iterrows(): if row["status"] == "canceled": continue rating = row["response"]["annotations"]["Preference"] if rating == "NaN": continue scores = row["response"]["responses"] if any(s["Preference"] == "" for s in scores): continue response = { "id": row["task_id"], "prompt": row["params"]["templateVariables"]["prompt"], "model_a": row["params"]["templateVariables"]["modela"], "model_b": row["params"]["templateVariables"]["modelb"], "response_a": row["params"]["templateVariables"]["response1"], "response_b": row["params"]["templateVariables"]["response2"], "rating": int(rating), "ratings": [np.array([s["Preference"] for s in scores], dtype=np.int32)], } if tie_allowed: response["win"] = ( "model_a" if response["rating"] < 4 else "model_b" if response["rating"] > 5 else "tie" ) else: response["win"] = "model_a" if response["rating"] < 5 else "model_b" responses.append(response) return pd.DataFrame(responses) def create_eval_df_for_gpt(df, tie_allowed): responses = [] for _, row in df.iterrows(): response = { "id": row["review_id"], "prompt": row["question"], "model_a": row["model1"], "model_b": row["model2"], "response_a": row["answer1"], "response_b": row["answer2"], "rating": row["score"][0], } if tie_allowed: response["win"] = ( "model_a" if response["rating"] < 4 else "model_b" if response["rating"] > 5 else "tie" ) else: response["win"] = "model_a" if response["rating"] < 5 else "model_b" responses.append(response) return pd.DataFrame(responses) # Compute the Elo rating for each model def compute_elo(df, k=32, scale=400, base=10, initial_rating=1000): rating = defaultdict(lambda: initial_rating) for _, model_a, model_b, win in df[["model_a", "model_b", "win"]].itertuples(): ra = rating[model_a] rb = rating[model_b] ea = 1 / (1 + base ** ((rb - ra) / scale)) eb = 1 / (1 + base ** ((ra - rb) / scale)) if win == "model_a": sa = 1 elif win == "model_b": sa = 0 elif win == "tie" or win == "tie (bothbad)": sa = 0.5 else: raise Exception(f"unexpected vote {win}") rating[model_a] += k * (sa - ea) rating[model_b] += k * (1 - sa - eb) return rating def convert_rating_from_float_to_int(df): return {model: int(rating) for model, rating in compute_elo(df).items()} def get_elo_results(df_instruct, df_code_instruct, tie_allowed): df_all = pd.concat([df_instruct, df_code_instruct]) df_gpt_4 = load_dataset( "gpt_4_evals/data/", split="train", revision="e007baaf6e505731c08a0bc1a833a1f8f8cb8846", ).to_pandas() dfs = [df_instruct, df_code_instruct, df_all] elo_ratings = [ convert_rating_from_float_to_int(create_eval_df(df, tie_allowed=tie_allowed)) for df in dfs ] gpt_4_elo_ratings = convert_rating_from_float_to_int( create_eval_df_for_gpt(df_gpt_4, tie_allowed=tie_allowed) ) elo_ratings.append(gpt_4_elo_ratings) results = [ EloEvalResult( model=model_name, gpt_4_all=elo_ratings[3][model_name], human_all=elo_ratings[2][model_name], human_instruct=elo_ratings[0][model_name], human_code_instruct=elo_ratings[1][model_name], tie_allowed=tie_allowed, ) for model_name in elo_ratings[0].keys() ] return results def get_elo_results_dicts(df_instruct, df_code_instruct, tie_allowed) -> List[Dict]: eval_results = get_elo_results(df_instruct, df_code_instruct, tie_allowed) return [r.to_dict() for r in eval_results] def get_elo_plots(df_instruct, df_code_instruct, tie_allowed): df_instruct = create_eval_df(df_instruct, tie_allowed=tie_allowed) df_code_instruct = create_eval_df(df_code_instruct, tie_allowed=tie_allowed) df_all = pd.concat([df_instruct, df_code_instruct]) game = df_all[["model_a", "model_b", "win"]] game_switch = switch_model_a_b(game) plot_1 = visualize_pairwise_win_fraction(game_switch, PLOT_1_TITLE) plot_2 = visualize_battle_count(game_switch, PLOT_2_TITLE) BOOTSTRAP_ROUNDS = 1000 if "bootstrap_elo_lu" not in globals(): bootstrap_elo_lu = get_bootstrap_result( game_switch, compute_elo, BOOTSTRAP_ROUNDS ) plot_3 = visualize_bootstrap_scores(bootstrap_elo_lu, PLOT_3_TITLE) plot_4 = visualize_rating_count(game, PLOT_4_TITLE) return plot_1, plot_2, plot_3, plot_4