Hannah
initial
cce0ed9
raw
history blame
9.14 kB
import json
from pathlib import Path
from typing import Dict, List, Optional
import numpy as np
import requests
from fastapi import FastAPI, HTTPException, Query
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
class LeaderboardModel(BaseModel):
model_name: str
type: str
model_link: Optional[str] = None
scores: Dict[str, float]
co2_cost: Optional[float] = None
class LeaderboardData(BaseModel):
models: List[LeaderboardModel]
updated_at: str
app = FastAPI(
title="LLM Leaderboard API",
description="API for serving Open LLM Leaderboard data",
version="1.0.0"
)
# Add CORS middleware to allow requests from your Gradio app
app.add_middleware(
CORSMiddleware,
allow_origins=["*"], # For production, specify your exact frontend URL
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
# Cache for leaderboard data
cached_data = None
cache_file = Path("leaderboard_cache.json")
def fetch_external_leaderboard_data(refresh: bool = False) -> Optional[Dict]:
"""
Fetch leaderboard data from external sources like HuggingFace.
Uses local cache if available and refresh is False.
"""
global cached_data
if not refresh and cached_data:
return cached_data
if not refresh and cache_file.exists():
try:
with open(cache_file) as f:
cached_data = json.load(f)
return cached_data
except:
pass # Fall back to fetching if cache read fails
try:
# Try different endpoints that might contain leaderboard data
endpoints = [
"https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard/raw/main/leaderboard_data.json",
"https://huggingface.co/api/spaces/HuggingFaceH4/open_llm_leaderboard/api/get_results",
]
for url in endpoints:
response = requests.get(url)
if response.status_code == 200:
data = response.json()
cached_data = data
with open(cache_file, "w") as f:
json.dump(data, f)
return data
# If all endpoints fail, return None
return None
except Exception as e:
print(f"Error fetching external leaderboard data: {e}")
return None
def generate_sample_data() -> Dict:
"""
Generate sample leaderboard data when external data can't be fetched.
"""
models = [
{"model_name": "meta-llama/llama-3-70b-instruct", "type": "open"},
{"model_name": "mistralai/Mistral-7B-Instruct-v0.3", "type": "open"},
{"model_name": "google/gemma-7b-it", "type": "open"},
{"model_name": "Qwen/Qwen2-7B-Instruct", "type": "open"},
{"model_name": "anthropic/claude-3-opus", "type": "closed", "external_link": "https://www.anthropic.com/claude"},
{"model_name": "OpenAI/gpt-4o", "type": "closed", "external_link": "https://openai.com/gpt-4"},
{"model_name": "01-ai/Yi-1.5-34B-Chat", "type": "open"},
{"model_name": "google/gemma-2b", "type": "open"},
{"model_name": "microsoft/phi-3-mini-4k-instruct", "type": "open"},
{"model_name": "microsoft/phi-3-mini-128k-instruct", "type": "open"},
{"model_name": "stabilityai/stable-beluga-7b", "type": "open"},
{"model_name": "togethercomputer/RedPajama-INCITE-7B-Instruct", "type": "open"},
{"model_name": "databricks/dbrx-instruct", "type": "closed", "external_link": "https://www.databricks.com/product/machine-learning/large-language-models"},
{"model_name": "mosaicml/mpt-7b-instruct", "type": "open"},
{"model_name": "01-ai/Yi-1.5-9B-Chat", "type": "open"},
{"model_name": "anthropic/claude-3-sonnet", "type": "closed", "external_link": "https://www.anthropic.com/claude"},
{"model_name": "cohere/command-r-plus", "type": "closed", "external_link": "https://cohere.com/models/command-r-plus"},
{"model_name": "meta-llama/llama-3-8b-instruct", "type": "open"}
]
np.random.seed(42) # For reproducibility
model_data = []
for model_info in models:
model_name = model_info["model_name"]
model_type = model_info["type"]
external_link = model_info.get("external_link", None)
# Generate random scores
average = round(np.random.uniform(40, 90), 2)
ifeval = round(np.random.uniform(30, 90), 2)
bbhi = round(np.random.uniform(40, 85), 2)
math = round(np.random.uniform(20, 80), 2)
gpqa = round(np.random.uniform(10, 70), 2)
mujb = round(np.random.uniform(10, 70), 2)
mmlu = round(np.random.uniform(40, 85), 2)
co2_cost = round(np.random.uniform(1, 100), 2)
# If it's an open model, it should have a link to Hugging Face
model_link = None
if external_link:
model_link = external_link
elif "/" in model_name:
model_link = f"https://huggingface.co/{model_name}"
else:
model_link = f"https://huggingface.co/models?search={model_name}"
model_data.append({
"model_name": model_name,
"type": model_type,
"model_link": model_link,
"scores": {
"average": average,
"ifeval": ifeval,
"bbhi": bbhi,
"math": math,
"gpqa": gpqa,
"mujb": mujb,
"mmlu": mmlu
},
"co2_cost": co2_cost
})
# Sort by average score
model_data.sort(key=lambda x: x["scores"]["average"], reverse=True)
# Create the final data structure
from datetime import datetime
leaderboard_data = {
"models": model_data,
"updated_at": datetime.now().isoformat()
}
return leaderboard_data
@app.get("/")
def read_root():
return {"message": "Welcome to the LLM Leaderboard API"}
@app.get("/api/leaderboard", response_model=LeaderboardData)
def get_leaderboard(refresh: bool = Query(False, description="Force refresh data from source")):
"""
Get the full leaderboard data.
If refresh is True, force fetch from source instead of using cache.
"""
external_data = fetch_external_leaderboard_data(refresh=refresh)
if external_data:
# Process external data to match our expected format
try:
# Here you would transform the external data to match LeaderboardData model
# This is a simplified example - you'd need to adapt this to the actual structure
return external_data
except Exception as e:
print(f"Error processing external data: {e}")
# Fall back to sample data if external data can't be processed
return generate_sample_data()
@app.get("/api/models", response_model=List[str])
def get_models():
"""Get a list of all model names in the leaderboard"""
data = fetch_external_leaderboard_data() or generate_sample_data()
return [model["model_name"] for model in data["models"]]
@app.get("/api/model/{model_name}", response_model=LeaderboardModel)
def get_model_details(model_name: str):
"""Get detailed information about a specific model"""
data = fetch_external_leaderboard_data() or generate_sample_data()
for model in data["models"]:
if model["model_name"] == model_name:
return model
raise HTTPException(status_code=404, detail=f"Model {model_name} not found")
@app.get("/api/filters")
def get_filter_counts():
"""
Get counts for different filter categories to display in the UI.
This matches what's shown in the 'Quick Filters' section of the leaderboard.
"""
data = fetch_external_leaderboard_data() or generate_sample_data()
# Count models by different categories
edge_count = 0
consumer_count = 0
midrange_count = 0
gpu_rich_count = 0
official_count = 0
for model in data["models"]:
# Edge devices (typically small models)
if "scores" in model and model["scores"].get("average", 0) < 45:
edge_count += 1
# Consumer (moderate size/performance)
if "scores" in model and 45 <= model["scores"].get("average", 0) < 55:
consumer_count += 1
# Mid-range
if "scores" in model and 55 <= model["scores"].get("average", 0) < 65:
midrange_count += 1
# GPU-rich (high-end models)
if "scores" in model and model["scores"].get("average", 0) >= 65:
gpu_rich_count += 1
# Official providers
# This is just a placeholder logic - adapt to your actual criteria
if "/" not in model["model_name"] or model["model_name"].startswith("meta/") or model["model_name"].startswith("google/"):
official_count += 1
return {
"edge_devices": edge_count,
"consumers": consumer_count,
"midrange": midrange_count,
"gpu_rich": gpu_rich_count,
"official_providers": official_count
}
if __name__ == "__main__":
import uvicorn
uvicorn.run(app, host="0.0.0.0", port=8000)