crfm-helm 0.4.0__py3-none-any.whl → 0.5.10__py3-none-any.whl

helm/proxy/services/remote_service.py

@@ -5,11 +5,16 @@ import urllib.parse
 from dataclasses import asdict
 from typing import Any, List, Optional
 
+from helm.common.cache import CacheConfig
+from helm.common.cache_backend_config import BlackHoleCacheBackendConfig
 from helm.common.authentication import Authentication
+from helm.common.moderations_api_request import ModerationAPIRequest, ModerationAPIRequestResult
 from helm.common.critique_request import CritiqueRequest, CritiqueRequestResult
+from helm.common.nudity_check_request import NudityCheckRequest, NudityCheckResult
+from helm.common.file_upload_request import FileUploadRequest, FileUploadResult
 from helm.common.perspective_api_request import PerspectiveAPIRequest, PerspectiveAPIRequestResult
+from helm.common.clip_score_request import CLIPScoreRequest, CLIPScoreResult
 from helm.common.tokenization_request import (
-    WindowServiceInfo,
     TokenizationRequest,
     TokenizationRequestResult,
     DecodeRequestResult,
@@ -19,7 +24,7 @@ from helm.common.request import Request, RequestResult
 from dacite import from_dict
 from helm.proxy.accounts import Account
 from helm.proxy.query import Query, QueryResult
-from .service import Service, GeneralInfo
+from helm.proxy.services.service import Service, GeneralInfo
 
 
 class RemoteServiceError(Exception):
@@ -27,6 +32,8 @@ class RemoteServiceError(Exception):
 
 
 class RemoteService(Service):
+    NOT_SUPPORTED_ERROR: str = "Not supported through the remote service."
+
     def __init__(self, base_url):
         self.base_url: str = base_url
 
@@ -43,11 +50,6 @@ class RemoteService(Service):
         response = requests.get(f"{self.base_url}/api/general_info").json()
         return from_dict(GeneralInfo, response)
 
-    def get_window_service_info(self, model_name) -> WindowServiceInfo:
-        params = {"model_name": model_name}
-        response = requests.get(f"{self.base_url}/api/window_service_info?{urllib.parse.urlencode(params)}").json()
-        return from_dict(WindowServiceInfo, response)
-
     def expand_query(self, query: Query) -> QueryResult:
         params = asdict(query)
         response = requests.get(f"{self.base_url}/api/query?{urllib.parse.urlencode(params)}").json()
@@ -84,6 +86,15 @@ class RemoteService(Service):
         RemoteService._check_response(response, request_json)
         return from_dict(DecodeRequestResult, response)
 
+    def upload(self, auth: Authentication, request: FileUploadRequest) -> FileUploadResult:
+        raise NotImplementedError(self.NOT_SUPPORTED_ERROR)
+
+    def check_nudity(self, auth: Authentication, request: NudityCheckRequest) -> NudityCheckResult:
+        raise NotImplementedError(self.NOT_SUPPORTED_ERROR)
+
+    def compute_clip_score(self, auth: Authentication, request: CLIPScoreRequest) -> CLIPScoreResult:
+        raise NotImplementedError(self.NOT_SUPPORTED_ERROR)
+
     def get_toxicity_scores(self, auth: Authentication, request: PerspectiveAPIRequest) -> PerspectiveAPIRequestResult:
         request_json: str = json.dumps(asdict(request))
         params = {
@@ -94,6 +105,16 @@ class RemoteService(Service):
         RemoteService._check_response(response, request_json)
         return from_dict(PerspectiveAPIRequestResult, response)
 
+    def get_moderation_results(self, auth: Authentication, request: ModerationAPIRequest) -> ModerationAPIRequestResult:
+        request_json: str = json.dumps(asdict(request))
+        params = {
+            "auth": json.dumps(asdict(auth)),
+            "request": request_json,
+        }
+        response = requests.get(f"{self.base_url}/api/moderation?{urllib.parse.urlencode(params)}").json()
+        RemoteService._check_response(response, request_json)
+        return from_dict(ModerationAPIRequestResult, response)
+
     def make_critique_request(self, auth: Authentication, request: CritiqueRequest) -> CritiqueRequestResult:
         raise NotImplementedError("make_critique_request is not supported by RemoteServer")
 
@@ -153,6 +174,10 @@ class RemoteService(Service):
             # A ConnectionError is expected when shutting down the server.
             pass
 
+    def get_cache_config(self, shard_name: str) -> CacheConfig:
+        """Returns a CacheConfig"""
+        return BlackHoleCacheBackendConfig().get_cache_config(shard_name)
+
 
 def add_service_args(parser: argparse.ArgumentParser):
     """Add command-line arguments to enable command-line utilities to specify how to connect to a remote server."""

helm/proxy/services/server_service.py

@@ -1,15 +1,18 @@
-import dataclasses
 import os
 import signal
-from typing import List, Optional
-from helm.common.cache_utils import build_cache_config
+from typing import List
 
+from helm.common.local_context import LocalContext
+from helm.common.cache import CacheConfig
+from helm.common.cache_backend_config import CacheBackendConfig, BlackHoleCacheBackendConfig
 from helm.common.critique_request import CritiqueRequest, CritiqueRequestResult
 from helm.common.authentication import Authentication
-from helm.common.general import ensure_directory_exists, parse_hocon, get_credentials
+from helm.common.moderations_api_request import ModerationAPIRequest, ModerationAPIRequestResult
+from helm.common.clip_score_request import CLIPScoreRequest, CLIPScoreResult
+from helm.common.nudity_check_request import NudityCheckRequest, NudityCheckResult
+from helm.common.file_upload_request import FileUploadRequest, FileUploadResult
 from helm.common.perspective_api_request import PerspectiveAPIRequest, PerspectiveAPIRequestResult
 from helm.common.tokenization_request import (
-    WindowServiceInfo,
     TokenizationRequest,
     TokenizationRequestResult,
     DecodeRequest,
@@ -18,24 +21,13 @@ from helm.common.tokenization_request import (
 from helm.common.request import Request, RequestResult
 from helm.common.hierarchical_logger import hlog
 from helm.proxy.accounts import Accounts, Account
-from helm.proxy.clients.auto_client import AutoClient
-from helm.proxy.clients.toxicity_classifier_client import ToxicityClassifierClient
-from helm.proxy.example_queries import example_queries
-from helm.benchmark.model_metadata_registry import ALL_MODELS_METADATA
 from helm.benchmark.model_deployment_registry import get_model_deployment_host_organization
 from helm.proxy.query import Query, QueryResult
-from helm.proxy.retry import retry_request
 from helm.proxy.token_counters.auto_token_counter import AutoTokenCounter
-from helm.proxy.tokenizers.auto_tokenizer import AutoTokenizer
-from helm.proxy.tokenizers.huggingface_tokenizer import HuggingFaceTokenizer
-from .service import (
+from helm.proxy.services.service import (
     Service,
-    CACHE_DIR,
     ACCOUNTS_FILE,
     GeneralInfo,
-    VERSION,
-    expand_environments,
-    synthesize_request,
 )
 
 
@@ -44,52 +36,46 @@ class ServerService(Service):
     Main class that supports various functionality for the server.
     """
 
-    def __init__(self, base_path: str = "prod_env", root_mode=False, mongo_uri: str = ""):
-        credentials = get_credentials(base_path)
-        cache_path = os.path.join(base_path, CACHE_DIR)
-        ensure_directory_exists(cache_path)
+    def __init__(
+        self,
+        base_path: str = "prod_env",
+        root_mode: bool = False,
+        cache_backend_config: CacheBackendConfig = BlackHoleCacheBackendConfig(),
+    ):
         accounts_path = os.path.join(base_path, ACCOUNTS_FILE)
 
-        self.client = AutoClient(credentials, cache_path, mongo_uri)
-        self.tokenizer = AutoTokenizer(credentials, cache_path, mongo_uri)
-        cache_config = build_cache_config(cache_path, mongo_uri, "huggingface")
-        self.token_counter = AutoTokenCounter(HuggingFaceTokenizer(cache_config=cache_config))
+        self.context = LocalContext(base_path, cache_backend_config)
+        self.token_counter = AutoTokenCounter(self.context.tokenizer)
         self.accounts = Accounts(accounts_path, root_mode=root_mode)
-        # Lazily instantiated by get_toxicity_scores()
-        self.toxicity_classifier_client: Optional[ToxicityClassifierClient] = None
 
     def get_general_info(self) -> GeneralInfo:
-        # Can't send release_dates in ModelMetadata bacause dates cannot be round-tripped to and from JSON easily.
-        # TODO(#2158): Either fix this or delete get_general_info.
-        all_models = [dataclasses.replace(model_metadata, release_date=None) for model_metadata in ALL_MODELS_METADATA]
-        return GeneralInfo(version=VERSION, example_queries=example_queries, all_models=all_models)
-
-    def get_window_service_info(self, model_name) -> WindowServiceInfo:
-        # The import statement is placed here to avoid two problems, please refer to the link for details
-        # https://github.com/stanford-crfm/helm/pull/1430#discussion_r1156686624
-        from helm.benchmark.window_services.tokenizer_service import TokenizerService
-        from helm.benchmark.window_services.window_service_factory import WindowServiceFactory
-
-        token_service = TokenizerService(self, Authentication(""))
-        window_service = WindowServiceFactory.get_window_service(model_name, token_service)
-        return WindowServiceInfo(
-            tokenizer_name=window_service.tokenizer_name,
-            max_sequence_length=window_service.max_sequence_length,
-            max_request_length=window_service.max_request_length,
-            end_of_text_token=window_service.end_of_text_token,
-            prefix_token=window_service.prefix_token,
-        )
+        return self.context.get_general_info()
 
     def expand_query(self, query: Query) -> QueryResult:
-        """Turn the `query` into requests."""
-        prompt = query.prompt
-        settings = query.settings
-        environments = parse_hocon(query.environments)
-        requests = []
-        for environment in expand_environments(environments):
-            request = synthesize_request(prompt, settings, environment)
-            requests.append(request)
-        return QueryResult(requests=requests)
+        return self.context.expand_query(query)
+
+    def _get_model_group_for_model_deployment(self, model_deployment: str) -> str:
+        if model_deployment.startswith("openai/"):
+            if model_deployment.startswith("openai/code-"):
+                return "codex"
+            elif model_deployment.startswith("openai/dall-e-"):
+                return "dall_e"
+            elif model_deployment.startswith("openai/gpt-4"):
+                return "gpt4"
+            elif model_deployment.startswith("openai/gpt-3"):
+                return "gpt3"
+            elif (
+                model_deployment.startswith("openai/o1")
+                or model_deployment.startswith("openai/o3")
+                or model_deployment.startswith("openai/o4")
+            ):
+                return "o1"
+            else:
+                return "openai"
+        elif model_deployment.startswith("ai21/"):
+            return "jurassic"
+        else:
+            return get_model_deployment_host_organization(model_deployment)
 
     def make_request(self, auth: Authentication, request: Request) -> RequestResult:
         """Actually make a request to an API."""
@@ -98,44 +84,57 @@ class ServerService(Service):
         #       https://github.com/stanford-crfm/benchmarking/issues/56
 
         self.accounts.authenticate(auth)
-        host_organization: str = get_model_deployment_host_organization(request.model_deployment)
+        model_group: str = self._get_model_group_for_model_deployment(request.model_deployment)
         # Make sure we can use
-        self.accounts.check_can_use(auth.api_key, host_organization)
+        self.accounts.check_can_use(auth.api_key, model_group)
 
         # Use!
-        request_result: RequestResult = self.client.make_request(request)
+        request_result: RequestResult = self.context.make_request(request)
 
         # Only deduct if not cached
         if not request_result.cached:
             # Count the number of tokens used
             count: int = self.token_counter.count_tokens(request, request_result.completions)
-            self.accounts.use(auth.api_key, host_organization, count)
+            self.accounts.use(auth.api_key, model_group, count)
 
         return request_result
 
     def tokenize(self, auth: Authentication, request: TokenizationRequest) -> TokenizationRequestResult:
         """Tokenize via an API."""
         self.accounts.authenticate(auth)
-        return self.tokenizer.tokenize(request)
+        return self.context.tokenize(request)
 
     def decode(self, auth: Authentication, request: DecodeRequest) -> DecodeRequestResult:
         """Decodes to text."""
         self.accounts.authenticate(auth)
-        return self.tokenizer.decode(request)
+        return self.context.decode(request)
+
+    def upload(self, auth: Authentication, request: FileUploadRequest) -> FileUploadResult:
+        """Uploads a file to external storage."""
+        self.accounts.authenticate(auth)
+        return self.context.upload(request)
+
+    def check_nudity(self, auth: Authentication, request: NudityCheckRequest) -> NudityCheckResult:
+        """Check for nudity."""
+        self.accounts.authenticate(auth)
+        return self.context.check_nudity(request)
+
+    def compute_clip_score(self, auth: Authentication, request: CLIPScoreRequest) -> CLIPScoreResult:
+        """Computes CLIPScore for a given caption and image."""
+        self.accounts.authenticate(auth)
+        return self.context.compute_clip_score(request)
 
     def get_toxicity_scores(self, auth: Authentication, request: PerspectiveAPIRequest) -> PerspectiveAPIRequestResult:
-        @retry_request
-        def get_toxicity_scores_with_retry(request: PerspectiveAPIRequest) -> PerspectiveAPIRequestResult:
-            if not self.toxicity_classifier_client:
-                self.toxicity_classifier_client = self.client.get_toxicity_classifier_client()
-            return self.toxicity_classifier_client.get_toxicity_scores(request)
+        self.accounts.authenticate(auth)
+        return self.context.get_toxicity_scores(request)
 
+    def get_moderation_results(self, auth: Authentication, request: ModerationAPIRequest) -> ModerationAPIRequestResult:
         self.accounts.authenticate(auth)
-        return get_toxicity_scores_with_retry(request)
+        return self.context.get_moderation_results(request)
 
     def make_critique_request(self, auth: Authentication, request: CritiqueRequest) -> CritiqueRequestResult:
         self.accounts.authenticate(auth)
-        return self.client.get_critique_client().make_critique_request(request)
+        return self.context.make_critique_request(request)
 
     def create_account(self, auth: Authentication) -> Account:
         """Creates a new account."""
@@ -168,3 +167,6 @@ class ServerService(Service):
         hlog(f"Shutting down server by killing its own process {pid}...")
         os.kill(pid, signal.SIGTERM)
         hlog("Done.")
+
+    def get_cache_config(self, shard_name: str) -> CacheConfig:
+        return self.context.get_cache_config(shard_name)

helm/proxy/services/service.py

@@ -5,9 +5,12 @@ from typing import Dict, List, Tuple, Any
 
 from helm.common.general import parse_hocon
 from helm.common.critique_request import CritiqueRequest, CritiqueRequestResult
+from helm.common.clip_score_request import CLIPScoreRequest, CLIPScoreResult
+from helm.common.file_upload_request import FileUploadResult, FileUploadRequest
+from helm.common.nudity_check_request import NudityCheckRequest, NudityCheckResult
 from helm.common.perspective_api_request import PerspectiveAPIRequestResult, PerspectiveAPIRequest
+from helm.common.moderations_api_request import ModerationAPIRequest, ModerationAPIRequestResult
 from helm.common.tokenization_request import (
-    WindowServiceInfo,
     TokenizationRequest,
     TokenizationRequestResult,
     DecodeRequest,
@@ -17,6 +20,7 @@ from helm.common.request import Request, RequestResult
 from helm.benchmark.model_metadata_registry import ModelMetadata
 from helm.proxy.query import Query, QueryResult
 from helm.proxy.accounts import Authentication, Account
+from helm.common.cache import CacheConfig
 
 VERSION = "1.0"
 ACCOUNTS_FILE = "accounts.sqlite"
@@ -80,11 +84,6 @@ class Service(ABC):
         """Get general info."""
         pass
 
-    @abstractmethod
-    def get_window_service_info(self, model_name: str) -> WindowServiceInfo:
-        """Get window service info."""
-        pass
-
     @abstractmethod
     def expand_query(self, query: Query) -> QueryResult:
         """Turn the `query` into requests."""
@@ -105,11 +104,31 @@ class Service(ABC):
         """Decodes to text."""
         pass
 
+    @abstractmethod
+    def upload(self, auth: Authentication, request: FileUploadRequest) -> FileUploadResult:
+        """Uploads a file to external storage."""
+        pass
+
+    @abstractmethod
+    def check_nudity(self, auth: Authentication, request: NudityCheckRequest) -> NudityCheckResult:
+        """Check for nudity for a batch of images."""
+        pass
+
+    @abstractmethod
+    def compute_clip_score(self, auth: Authentication, request: CLIPScoreRequest) -> CLIPScoreResult:
+        """Computes CLIPScore for a given caption and image."""
+        pass
+
     @abstractmethod
     def get_toxicity_scores(self, auth: Authentication, request: PerspectiveAPIRequest) -> PerspectiveAPIRequestResult:
         """Get toxicity scores for a batch of text."""
         pass
 
+    @abstractmethod
+    def get_moderation_results(self, auth: Authentication, request: ModerationAPIRequest) -> ModerationAPIRequestResult:
+        """Get OpenAI's moderation results for some text."""
+        pass
+
     @abstractmethod
     def make_critique_request(self, auth: Authentication, request: CritiqueRequest) -> CritiqueRequestResult:
         """Get responses to a critique request."""
@@ -149,3 +168,8 @@ class Service(ABC):
     def shutdown(self, auth: Authentication):
         """Shutdown server."""
         pass
+
+    @abstractmethod
+    def get_cache_config(self, shard_name: str) -> CacheConfig:
+        """Returns a CacheConfig"""
+        pass

helm/proxy/services/test_remote_service.py

@@ -17,9 +17,9 @@ from sqlitedict import SqliteDict
 from helm.common.authentication import Authentication
 from helm.common.request import Request, RequestResult
 from helm.common.tokenization_request import TokenizationRequest, TokenizationRequestResult
-from helm.proxy.accounts import Account
-from .remote_service import RemoteService
-from .service import ACCOUNTS_FILE
+from helm.proxy.accounts import Account, set_default_quotas
+from helm.proxy.services.remote_service import RemoteService
+from helm.proxy.services.service import ACCOUNTS_FILE
 
 
 @dataclass(frozen=True)
@@ -55,6 +55,7 @@ class TestRemoteServerService:
 
             with SqliteDict(os.path.join(path, ACCOUNTS_FILE)) as cache:
                 account: Account = Account(TestRemoteServerService._ADMIN_API_KEY, is_admin=True)
+                set_default_quotas(account)
                 cache[TestRemoteServerService._ADMIN_API_KEY] = asdict(account)
                 cache.commit()
             return path
@@ -126,9 +127,9 @@ class TestRemoteServerService:
         assert response.success
 
     def test_tokenize(self):
-        request = TokenizationRequest(text="1 2 3", tokenizer="simple/model1")
+        request = TokenizationRequest(text="1 2 3", tokenizer="simple/tokenizer1")
         response: TokenizationRequestResult = self.service.tokenize(self.auth, request)
-        assert [token.value for token in response.tokens] == ["1", "2", "3"]
+        assert [token.value for token in response.tokens] == ["1", " ", "2", " ", "3"]
 
     def test_make_request_plus_sign(self):
         # Ensure + in prompt doesn't get replaced by a blank space

helm/proxy/services/test_service.py

@@ -8,7 +8,7 @@ from helm.common.authentication import Authentication
 from helm.common.request import Request
 from helm.proxy.accounts import AuthenticationError, Accounts
 from helm.proxy.query import Query
-from .server_service import ServerService
+from helm.proxy.services.server_service import ServerService
 
 
 def get_authentication():
@@ -197,18 +197,6 @@ def helper_prod_test_service(request: Request, expected_text: str):
         # Consistency of log probs
         assert completion.logprob == sum(token.logprob for token in completion.tokens)
 
-        for token in completion.tokens[1:]:
-            assert len(token.top_logprobs) == request.top_k_per_token
-
-            # If generated token was one of the top, make sure has the right probability
-            if token.text in token.top_logprobs:
-                assert token.logprob == token.top_logprobs[token.text]
-
-            # If temperature = 0, then make sure we're getting the top probability token
-            if request.temperature == 0:
-                assert token.text in token.top_logprobs
-                assert token.logprob == max(token.top_logprobs.values())
-
     # Make sure we get the expected_text in one of the completions
     assert any(completion.text == expected_text for completion in result.completions)
 

helm/proxy/static/help.html

@@ -0,0 +1,99 @@
+<!DOCTYPE html>
+<html lang="en">
+	<head>
+		<title>CRFM Models API</title>
+		<meta charset="utf-8">
+		<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
+		<link rel="stylesheet" href="https://maxcdn.bootstrapcdn.com/bootstrap/4.0.0-beta/css/bootstrap.min.css" integrity="sha384-/Y6pD6FV/Vv2HJnA6t+vslU6fwYXjCFtcEpHbNJ0lyAFsXTsjBbfaDjzALeQsN6M" crossorigin="anonymous">
+		<link rel="stylesheet" type="text/css" href="index.css">
+	</head>
+
+	<body>
+		<div class="container-fluid">
+      <nav class="navbar navbar-expand-sm navbar-light bg-faded">
+        <button class="navbar-toggler" type="button" data-toggle="collapse" data-target="#nav-content" aria-controls="nav-content" aria-expanded="false" aria-label="Toggle navigation">
+          <span class="navbar-toggler-icon"></span>
+        </button>
+
+        <div class="collapse navbar-collapse" id="nav-content">
+          <ul class="navbar-nav">
+            <li class="nav-item"><a class="nav-link" href="index.html">Query</a></li>
+            <li class="nav-item"><a class="nav-link active" href="help.html">Help</a></li>
+            <li class="nav-item" id="loginInfo"></li>
+          </ul>
+        </div>
+      </nav>
+
+      <div class="row">
+        <div class="col-sm-12">
+          <p>
+            This web interface and API allows you to experiment with multiple large language models using a unified interface.
+            To log in, you need an API key.
+          </p>
+          <h2>Query</h2>
+          <ul>
+            <li>In the web interface, you can enter a <b>query</b>, which consists of the following components:
+              <ul>
+                <li><b>prompt</b>, which is what text we want to feed into the language model.  The prompt can have variables (e.g., <tt>${name}</tt>) which are filled in later.</li>
+                <li><b>settings</b>, which configures how we're going to call the backend API (HOCON format):
+                  <ul>
+                    <li><tt>model</tt>: which model to query; options are:
+                      <div id="help-models" />
+                    </li>
+                    <li><tt>temperature</tt>: a non-negative number determining amount of stochasticity (e.g., <tt>1</tt> is sampling from the model, <tt>0</tt> is returning the maximum probability output)</li>
+                    <li><tt>num_completions</tt>: number of completions (sequences, independent sampled) to return</li>
+                    <li><tt>top_k_per_token</tt>: number of candidates per token position in each completion</li>
+                    <li><tt>max_tokens</tt>: maximum number of tokens before generation stops</li>
+                    <li><tt>stop_sequences</tt>: list of strings that will stop generation (e.g., <tt>'.'</tt> or <tt>'\n'</tt>)</li>
+                    <li><tt>echo_prompt</tt>: Whether to include the prompt as a prefix of the completion</li>
+                    <li><tt>top_p</tt>: An alternative to sampling with temperature, called nucleus sampling, where the model considers the results of the tokens with top_p probability mass. So 0.1 means only the tokens comprising the top 10% probability mass are considered.</li>
+                    <li><tt>presence_penalty</tt>: Number between -2.0 and 2.0. Positive values penalize new tokens based on whether they appear in the text so far, increasing the model's likelihood to talk about new topics. (OpenAI only)</li>
+                    <li><tt>frequency_penalty</tt>: Number between -2.0 and 2.0. Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim. (OpenAI only)</li>
+                  </ul>
+                  Settings can also have variables in them (e.g., <tt>temperature</tt>).
+                </li>
+                <li><b>environments</b>, which specifies for each variable, a list of values (HOCON format).
+              </ul>
+            </li>
+            <li>
+              When the query is submitted,
+              we consider all possible assignments of values to variables.
+              For example:
+              <ul>
+                <li>environments has <tt>name: [Boston, New York]</tt> and <tt>temperature: [0, 1]</tt></li>
+                <li>prompt is <tt>${name} is a</tt></li>
+                <li>settings is <tt>temperature: ${temperature}</tt></li>
+              </ul>
+              This gives rise to 4 <b>requests</b>:
+              <ul>
+                <li>prompt: <tt>Boston is a</tt>, temperature: <tt>0</tt></li>
+                <li>prompt: <tt>Boston is a</tt>, temperature: <tt>1</tt></li>
+                <li>prompt: <tt>New York is a</tt>, temperature: <tt>0</tt></li>
+                <li>prompt: <tt>New York is a</tt>, temperature: <tt>1</tt></li>
+              </ul>
+            </li>
+          </ul>
+
+          <h2><a name="quotas">Quotas</a></h2>
+          <p>
+            For each model group (e.g., <tt>gpt3</tt>) and time granularity
+            (e.g., <tt>daily</tt>, <tt>monthly</tt>, <tt>total</tt>), you are
+            given a quota of a certain number of tokens.
+            Once you go over that number, you won't be able to use the API.
+            However, note that requests that have already been made (by you or
+            anyone) that are cached are not counted towards your quota.
+            For example, if your daily quota for <tt>gpt3</tt> is <tt>10000</tt>,
+            that means each day, you will get 10000 tokens.
+          </p>
+        </div>
+      </div>
+		</div>
+
+		<script src="https://code.jquery.com/jquery-3.2.1.min.js"></script>
+		<script src="https://cdnjs.cloudflare.com/ajax/libs/popper.js/1.11.0/umd/popper.min.js" integrity="sha384-b/U6ypiBEHpOf/4+1nzFpr53nxSS+GLCkfwBdFNTxtclqqenISfwAzpKaMNFNmj4" crossorigin="anonymous"></script>
+		<script src="https://maxcdn.bootstrapcdn.com/bootstrap/4.0.0-beta/js/bootstrap.min.js" integrity="sha384-h0AbiXch4ZDo7tp9hKZ4TsHbi047NrKGLO3SEJAg45jXxnGIfYzk4Si90RDIqNm1" crossorigin="anonymous"></script>
+    <script src="https://cdnjs.cloudflare.com/ajax/libs/jquery.tablesorter/2.29.0/js/jquery.tablesorter.min.js"></script>
+		<script src="general.js"></script>
+		<script src="index.js"></script>
+	</body>
+</html>

helm/proxy/static/index.css

@@ -0,0 +1,61 @@
+.warning {
+  color: red;
+}
+
+.active {
+  background-color: #f0f0f0;
+}
+
+.examples-block {
+  margin-left: 120px;
+  margin-top: 10px;
+  margin-bottom: 10px;
+}
+
+.request {
+  background-color: #f0f0f0;
+  margin-top: 20px;
+  margin-bottom: 20px;
+}
+
+.completion {
+  border: solid 1px;
+  border-color: #c0c0c0;
+  padding: 3px;
+  margin-top: 5px;
+  margin-bottom: 5px;
+}
+
+.completion:hover {
+  background-color: #e8e8e8;
+}
+
+.metadata {
+  float: right;
+  font-style: italic;
+}
+
+.thinking {
+  font-style: italic;
+}
+
+.token:hover {
+  background-color: lightgreen;
+}
+
+.block {
+  margin: 15px;
+  padding: 15px;
+  border: solid 1px;
+  border-color: lightgray;
+}
+
+td {
+  text-align: left;
+  vertical-align: top;
+}
+
+.help-icon {
+  margin-left: 3px;
+  margin-right: 3px;
+}

helm/proxy/static/index.html

@@ -0,0 +1,40 @@
+<!DOCTYPE html>
+<html lang="en">
+	<head>
+		<title>CRFM Models API</title>
+		<meta charset="utf-8">
+		<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
+		<link rel="stylesheet" href="https://maxcdn.bootstrapcdn.com/bootstrap/4.0.0-beta/css/bootstrap.min.css" integrity="sha384-/Y6pD6FV/Vv2HJnA6t+vslU6fwYXjCFtcEpHbNJ0lyAFsXTsjBbfaDjzALeQsN6M" crossorigin="anonymous">
+		<link rel="stylesheet" type="text/css" href="index.css">
+	</head>
+
+	<body>
+		<div class="container-fluid">
+      <nav class="navbar navbar-expand-sm navbar-light bg-faded">
+        <button class="navbar-toggler" type="button" data-toggle="collapse" data-target="#nav-content" aria-controls="nav-content" aria-expanded="false" aria-label="Toggle navigation">
+          <span class="navbar-toggler-icon"></span>
+        </button>
+
+        <div class="collapse navbar-collapse" id="nav-content">
+          <ul class="navbar-nav">
+            <li class="nav-item"><a class="nav-link active" href="index.html">Query</a></li>
+            <li class="nav-item"><a class="nav-link" href="help.html">Help</a></li>
+            <li class="nav-item" id="loginInfo"></li>
+          </ul>
+        </div>
+      </nav>
+
+      <div class="row">
+        <div class="col-sm-12" id="main">
+        </div>
+      </div>
+		</div>
+
+		<script src="https://code.jquery.com/jquery-3.2.1.min.js"></script>
+		<script src="https://cdnjs.cloudflare.com/ajax/libs/popper.js/1.11.0/umd/popper.min.js" integrity="sha384-b/U6ypiBEHpOf/4+1nzFpr53nxSS+GLCkfwBdFNTxtclqqenISfwAzpKaMNFNmj4" crossorigin="anonymous"></script>
+		<script src="https://maxcdn.bootstrapcdn.com/bootstrap/4.0.0-beta/js/bootstrap.min.js" integrity="sha384-h0AbiXch4ZDo7tp9hKZ4TsHbi047NrKGLO3SEJAg45jXxnGIfYzk4Si90RDIqNm1" crossorigin="anonymous"></script>
+    <script src="https://cdnjs.cloudflare.com/ajax/libs/jquery.tablesorter/2.29.0/js/jquery.tablesorter.min.js"></script>
+		<script src="general.js"></script>
+		<script src="index.js"></script>
+	</body>
+</html>