evals/models.py

import re
from datetime import date
from os import getenv

import pandas as pd
from aiolimiter import AsyncLimiter
from dotenv import load_dotenv
from google.cloud import translate_v2 as translate
from huggingface_hub import AsyncInferenceClient, HfApi
from joblib.memory import Memory
from openai import AsyncOpenAI, BadRequestError
from requests import HTTPError, get

# for development purposes, all languages will be evaluated on the fast models
# and only a sample of languages will be evaluated on all models
important_models = [
    "meta-llama/llama-4-maverick",  # 0.6$
    "meta-llama/llama-3.3-70b-instruct",  # 0.3$
    "meta-llama/llama-3.1-70b-instruct",  # 0.3$
    "meta-llama/llama-3-70b-instruct",  # 0.4$
    # "meta-llama/llama-2-70b-chat", # 0.9$; not properly supported by OpenRouter
    "openai/gpt-5",
    "openai/gpt-5-mini",
    "openai/gpt-5-nano",
    "openai/gpt-4.1",  # 8$
    "openai/gpt-4o",  # 10$
    "openai/gpt-3.5-turbo", # $1.50
    "openai/gpt-oss-120b",
    "anthropic/claude-sonnet-4.5",
    "anthropic/claude-haiku-4.5",
    "anthropic/claude-opus-4.1",  # 15$
    "anthropic/claude-sonnet-4",
    "anthropic/claude-3.7-sonnet",  # 15$
    "anthropic/claude-3.5-sonnet",
    "mistralai/mistral-small-3.2-24b-instruct",  # 0.3$
    "mistralai/mistral-medium-3.1",
    "mistralai/mistral-saba",  # 0.6$
    "mistralai/mistral-nemo",  # 0.08$
    "google/gemini-2.5-pro", # $10
    "google/gemini-2.5-flash",  # 0.6$
    "google/gemini-2.5-flash-lite",  # 0.3$
    "google/gemma-3-27b-it",  # 0.2$
    # "x-ai/grok-4", # $15
    "x-ai/grok-4-fast", 
    # "x-ai/grok-3", # $15
    "cohere/command-a",
    "qwen/qwen3-32b",
    "qwen/qwen3-235b-a22b",
    "qwen/qwen3-30b-a3b",  # 0.29$
    # "qwen/qwen-turbo", # 0.2$; recognizes "inappropriate content"
    # "qwen/qwq-32b",  # 0.2$
    # "qwen/qwen-2.5-72b-instruct",  # 0.39$
    # "qwen/qwen-2-72b-instruct",  # 0.9$
    "deepseek/deepseek-v3.2-exp",
    "microsoft/phi-4",  # 0.07$
    "amazon/nova-premier-v1", # 12.5$
    "amazon/nova-pro-v1",  # 0.09$
    "moonshotai/kimi-k2",  # 0.6$
    # "moonshotai/kimi-k2-thinking", # 2.5$
    "baidu/ernie-4.5-300b-a47b",
    # "baidu/ernie-4.5-21b-a3b-thinking",
    "z-ai/glm-4.6", # 1.75$
]

blocklist = [
    "google/gemini-2.5-pro-preview",
    # "google/gemini-2.5-pro",
    "google/gemini-2.5-flash-preview",
    "google/gemini-2.5-flash-lite-preview",
    "google/gemini-2.5-flash-preview-04-17",
    "google/gemini-2.5-flash-preview-05-20",
    "google/gemini-2.5-flash-lite-preview-06-17",
    "google/gemini-2.5-pro-preview-06-05",
    "google/gemini-2.5-pro-preview-05-06",
    "perplexity/sonar-deep-research",
    "perplexity/sonar-reasoning",
    "perplexity/sonar-reasoning-pro",
    "qwen/qwen3-vl-30b-a3b-thinking",
    "alpindale/goliath-120b"
]

transcription_models = [
    "elevenlabs/scribe_v1",
    "openai/whisper-large-v3",
    # "openai/whisper-small",
    # "facebook/seamless-m4t-v2-large",
]

cache = Memory(location=".cache", verbose=0).cache


@cache
def load_or_metadata(date: date):
    return get("https://openrouter.ai/api/frontend/models").json()["data"]


def get_or_metadata(permaslug):
    models = load_or_metadata(date.today())
    slugs = [
        m
        for m in models
        if (m["permaslug"] == permaslug or m["slug"] == permaslug)
        # ensure that a provider endpoint is available
        and m["endpoint"]
        # exclude free models
        # the problem is that free models typically have very high rate-limiting
        and not m["endpoint"]["is_free"]
        # exclude providers that train on user data
        # this is crucial since we are submitting benchmark data
        # make sure to additionally configure this in OpenRouter settings to avoid mistakes!
        and m["endpoint"]["provider_info"]["dataPolicy"]["training"] is False
    ]
    if len(slugs) == 0:
        print(f"no appropriate model (not free and no user data training) found for {permaslug}")
    return slugs[0] if len(slugs) >= 1 else None


@cache
def get_historical_popular_models(date: date):
    # date parameter is used for daily caching
    try:
        raw = get("https://openrouter.ai/rankings").text

        # Extract model data from rankingData using regex
        # Find all count and model_permaslug pairs in the data
        # Format: "count":number,"model_permaslug":"model/name"
        pattern = r"\\\"count\\\":([\d.]+).*?\\\"model_permaslug\\\":\\\"([^\\\"]+)\\\""
        matches = re.findall(pattern, raw)

        if matches:
            # Aggregate model counts
            model_counts = {}
            for count_str, model_slug in matches:
                count = float(count_str)
                if not model_slug.startswith("openrouter") and model_slug != "Others":
                    # Remove variant suffixes for aggregation
                    base_model = model_slug.split(":")[0]
                    model_counts[base_model] = model_counts.get(base_model, 0) + count

            # Sort by popularity and return top models
            sorted_models = sorted(
                model_counts.items(), key=lambda x: x[1], reverse=True
            )
            result = []
            for model_slug, count in sorted_models:
                result.append({"slug": model_slug, "count": int(count)})

            return result
        else:
            return []

    except Exception as e:
        return []


@cache
def get_current_popular_models(date: date):
    # date parameter is used for daily caching
    try:
        raw = get("https://openrouter.ai/rankings?view=day").text

        # Extract model data from daily rankings
        # Find all count and model_permaslug pairs in the daily data
        pattern = r"\\\"count\\\":([\d.]+).*?\\\"model_permaslug\\\":\\\"([^\\\"]+)\\\""
        matches = re.findall(pattern, raw)

        if matches:
            # Aggregate model counts
            model_counts = {}
            for count_str, model_slug in matches:
                count = float(count_str)
                if not model_slug.startswith("openrouter") and model_slug != "Others":
                    # Remove variant suffixes for aggregation
                    base_model = model_slug.split(":")[0]
                    model_counts[base_model] = model_counts.get(base_model, 0) + count

            # Sort by popularity and return top models
            sorted_models = sorted(
                model_counts.items(), key=lambda x: x[1], reverse=True
            )
            result = []
            for model_slug, count in sorted_models:
                result.append({"slug": model_slug, "count": int(count)})

            return result
        else:
            return []

    except Exception as e:
        return []


def get_translation_models():
    return pd.DataFrame(
        [
            {
                "id": "google/translate-v2",
                "name": "Google Translate",
                "provider_name": "Google",
                "cost": 20.0,
                "train_on_prompts": False,  # they don't do it in the API
                "size": None,
                "type": "closed-source",
                "license": None,
                "tasks": ["translation_from", "translation_to"],
            }
        ]
    )


load_dotenv()
client = AsyncOpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=getenv("OPENROUTER_API_KEY"),
)

openrouter_rate_limit = AsyncLimiter(max_rate=20, time_period=1)
elevenlabs_rate_limit = AsyncLimiter(max_rate=2, time_period=1)
huggingface_rate_limit = AsyncLimiter(max_rate=5, time_period=1)
google_rate_limit = AsyncLimiter(max_rate=10, time_period=1)


@cache
async def complete(**kwargs) -> str | None:
    async with openrouter_rate_limit:
        try:
            response = await client.chat.completions.create(**kwargs)
        except BadRequestError as e:
            if "filtered" in e.message:
                return None
            raise e
    if not response.choices:
        raise Exception(response)
    return response.choices[0].message.content.strip()


translate_client = translate.Client()


def get_google_supported_languages():
    return [l["language"] for l in translate_client.get_languages()]


@cache
async def translate_google(text, source_language, target_language):
    async with google_rate_limit:
        response = translate_client.translate(
            text, source_language=source_language, target_language=target_language
        )
    return response["translatedText"]


# @cache
# async def transcribe_elevenlabs(path, model):
#     modelname = model.split("/")[-1]
#     client = AsyncElevenLabs(api_key=getenv("ELEVENLABS_API_KEY"))
#     async with elevenlabs_rate_limit:
#         with open(path, "rb") as file:
#             response = await client.speech_to_text.convert(
#                 model_id=modelname, file=file
#             )
#     return response.text


# @cache
# async def transcribe_huggingface(path, model):
#     client = AsyncInferenceClient(api_key=getenv("HUGGINGFACE_ACCESS_TOKEN"))
#     async with huggingface_rate_limit:
#         output = await client.automatic_speech_recognition(model=model, audio=path)
#     return output.text


# async def transcribe(path, model="elevenlabs/scribe_v1"):
#     provider, modelname = model.split("/")
#     match provider:
#         case "elevenlabs":
#             return await transcribe_elevenlabs(path, modelname)
#         case "openai" | "facebook":
#             return await transcribe_huggingface(path, model)
#         case _:
#             raise ValueError(f"Model {model} not supported")


api = HfApi()


@cache
def get_hf_metadata(row):
    # get metadata from the HuggingFace API
    empty = {
        "hf_id": None,
        "creation_date": None,
        "size": None,
        "type": "closed-source",
        "license": None,
    }
    if not row:
        return empty
    id = row["hf_slug"] or row["slug"].split(":")[0]
    if not id:
        return empty
    try:
        info = api.model_info(id)
        license = ""
        if (
            info.card_data
            and hasattr(info.card_data, "license")
            and info.card_data.license
        ):
            license = (
                info.card_data.license.replace("-", " ").replace("mit", "MIT").title()
            )
        return {
            "hf_id": info.id,
            "creation_date": info.created_at,
            "size": info.safetensors.total if info.safetensors else None,
            "type": "open-source",
            "license": license,
        }
    except HTTPError:
        return empty


def get_cost(row):
    try:
        cost = float(row["endpoint"]["pricing"]["completion"])
        return round(cost * 1_000_000, 2)
    except (TypeError, KeyError):
        return None


def get_training_policy(row):
    # get openrouter info whether the provider may train on prompts
    # (this needs to be thoroughly avoided for our benchmark prompts!)
    return row["endpoint"]["provider_info"]["dataPolicy"]["training"]


@cache
def load_models(date: date) -> pd.DataFrame:
    # popular_models = (
    #     get_historical_popular_models(date.today())[:20]
    #     + get_current_popular_models(date.today())[:10]
    # )
    popular_models = []
    popular_models = [m["slug"] for m in popular_models]
    all_model_candidates = set(important_models + popular_models) - set(blocklist)

    # Validate models exist on OpenRouter before including them
    valid_models = []

    for model_id in all_model_candidates:
        metadata = get_or_metadata(model_id)
        if metadata is not None:
            valid_models.append(model_id)

    models = pd.DataFrame(sorted(valid_models), columns=["id"])
    or_metadata = models["id"].apply(get_or_metadata)  # TODO this is double-doubled
    hf_metadata = or_metadata.apply(get_hf_metadata)
    creation_date_hf = pd.to_datetime(hf_metadata.str["creation_date"]).dt.date
    creation_date_or = pd.to_datetime(
        or_metadata.str["created_at"].str.split("T").str[0]
    ).dt.date

    models = models.assign(
        name=or_metadata.str["short_name"]
        .str.replace(" (free)", "")
        .str.replace(" (self-moderated)", "")
        .str.replace(r"\s*\([^)]*\)\s*$", "", regex=True),
        provider_name=or_metadata.str["name"].str.split(": ").str[0],
        # openrouter_metadata=or_metadata.astype(str),
        cost=or_metadata.apply(get_cost),
        train_on_prompts=or_metadata.apply(get_training_policy),
        hf_id=hf_metadata.str["hf_id"],
        size=hf_metadata.str["size"],
        type=hf_metadata.str["type"],
        license=hf_metadata.str["license"],
        creation_date=creation_date_hf.combine_first(creation_date_or),
    )
    models.to_json(
        "models_unfiltered.json", orient="records", indent=2, force_ascii=False
    )
    # Filter out expensive models to keep costs reasonable
    models = models[models["cost"] <= 15.0].reset_index(drop=True)
    models["tasks"] = [
        [
            "translation_from",
            "translation_to",
            "classification",
            "mmlu",
            "arc",
            "truthfulqa",
            "mgsm",
        ]
    ] * len(models)
    models = pd.concat([models, get_translation_models()])
    return models


models = load_models(date.today())