Add Command R and Command R+ models (#2548)

Co-authored-by: Yifan Mai <[email protected]>

setup.cfg

@@ -129,6 +129,9 @@ anthropic =
     anthropic~=0.17
     websocket-client~=1.3.2  # For legacy stanford-online-all-v4-s3
 
+cohere =
+    cohere~=5.3
+
 mistral =
     mistralai~=0.0.11
 
@@ -154,6 +157,7 @@ models =
     crfm-helm[allenai]
     crfm-helm[amazon]
     crfm-helm[anthropic]
+    crfm-helm[cohere]
     crfm-helm[google]
     crfm-helm[mistral]
     crfm-helm[openai]

src/helm/clients/cohere_client.py

@@ -1,8 +1,9 @@
 import json
 import requests
-from typing import List
+from typing import List, Optional, Sequence, TypedDict
 
 from helm.common.cache import CacheConfig
+from helm.common.optional_dependencies import handle_module_not_found_error
 from helm.common.request import (
     wrap_request_time,
     EMBEDDING_UNAVAILABLE_REQUEST_RESULT,
@@ -11,8 +12,13 @@
     GeneratedOutput,
     Token,
 )
-from .client import CachingClient, truncate_sequence
-from .cohere_utils import get_cohere_url, DEFAULT_COHERE_API_VERSION
+from helm.clients.client import CachingClient, truncate_sequence
+from helm.clients.cohere_utils import get_cohere_url, DEFAULT_COHERE_API_VERSION
+
+try:
+    import cohere
+except ModuleNotFoundError as e:
+    handle_module_not_found_error(e, ["cohere"])
 
 
 class CohereClient(CachingClient):
@@ -152,3 +158,92 @@ def do_it():
             completions=completions,
             embedding=[],
         )
+
+
+class CohereRawChatRequest(TypedDict):
+    message: str
+    model: Optional[str]
+    preamble: Optional[str]
+    chat_history: Optional[Sequence[cohere.ChatMessage]]
+    temperature: Optional[float]
+    max_tokens: Optional[int]
+    k: Optional[int]
+    p: Optional[float]
+    seed: Optional[float]
+    stop_sequences: Optional[Sequence[str]]
+    frequency_penalty: Optional[float]
+    presence_penalty: Optional[float]
+
+
+def convert_to_raw_chat_request(request: Request) -> CohereRawChatRequest:
+    # TODO: Support chat
+    model = request.model.replace("cohere/", "")
+    return {
+        "message": request.prompt,
+        "model": model,
+        "preamble": None,
+        "chat_history": None,
+        "temperature": request.temperature,
+        "max_tokens": request.max_tokens,
+        "k": request.top_k_per_token,
+        "p": request.top_p,
+        "stop_sequences": request.stop_sequences,
+        "seed": float(request.random) if request.random is not None else None,
+        "frequency_penalty": request.frequency_penalty,
+        "presence_penalty": request.presence_penalty,
+    }
+
+
+class CohereChatClient(CachingClient):
+    """
+    Leverages the chat endpoint: https://docs.cohere.com/reference/chat
+
+    Cohere models will only support chat soon: https://docs.cohere.com/docs/migrating-from-cogenerate-to-cochat
+    """
+
+    def __init__(self, api_key: str, cache_config: CacheConfig):
+        super().__init__(cache_config=cache_config)
+        self.client = cohere.Client(api_key=api_key)
+
+    def make_request(self, request: Request) -> RequestResult:
+        if request.embedding:
+            return EMBEDDING_UNAVAILABLE_REQUEST_RESULT
+        # TODO: Support multiple completions
+        assert request.num_completions == 1, "CohereChatClient only supports num_completions=1"
+        # TODO: Support messages
+        assert not request.messages, "CohereChatClient currently does not support the messages API"
+
+        raw_request: CohereRawChatRequest = convert_to_raw_chat_request(request)
+
+        try:
+
+            def do_it():
+                """
+                Send the request to the Cohere Chat API. Responses will be structured like this:
+                cohere.Chat {
+                    message: What's up?
+                    text: Hey there! How's it going? I'm doing well, thank you for asking 😊.
+                    ...
+                }
+                """
+                raw_response = self.client.chat(**raw_request).dict()
+                assert "text" in raw_response, f"Response does not contain text: {raw_response}"
+                return raw_response
+
+            response, cached = self.cache.get(raw_request, wrap_request_time(do_it))
+        except (requests.exceptions.RequestException, AssertionError) as e:
+            error: str = f"CohereClient error: {e}"
+            return RequestResult(success=False, cached=False, error=error, completions=[], embedding=[])
+
+        completions: List[GeneratedOutput] = []
+        completion: GeneratedOutput = GeneratedOutput(text=response["text"], logprob=0.0, tokens=[])
+        completions.append(completion)
+
+        return RequestResult(
+            success=True,
+            cached=cached,
+            request_time=response["request_time"],
+            request_datetime=response["request_datetime"],
+            completions=completions,
+            embedding=[],
+        )

src/helm/config/model_deployments.yaml

@@ -325,6 +325,25 @@ model_deployments:
     window_service_spec:
       class_name: "helm.benchmark.window_services.cohere_window_service.CohereWindowService"
 
+  - name: cohere/command-r
+    model_name: cohere/command-r
+    tokenizer_name: cohere/c4ai-command-r-v01
+    max_sequence_length: 128000
+    max_request_length: 128000
+    client_spec:
+      class_name: "helm.clients.cohere_client.CohereChatClient"
+
+  - name: cohere/command-r-plus
+    model_name: cohere/command-r-plus
+    tokenizer_name: cohere/c4ai-command-r-plus
+    # "We have a known issue where prompts between 112K - 128K in length
+    # result in bad generations."
+    # Source: https://docs.cohere.com/docs/command-r-plus
+    max_sequence_length: 110000
+    max_request_length: 110000
+    client_spec:
+      class_name: "helm.clients.cohere_client.CohereChatClient"
+
   # Craiyon
 
   - name: craiyon/dalle-mini

src/helm/config/model_metadata.yaml

@@ -468,7 +468,25 @@ models:
     creator_organization_name: Cohere
     access: limited
     release_date: 2023-09-29
-    tags: [TEXT_MODEL_TAG, FULL_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
+    tags: [TEXT_MODEL_TAG, PARTIAL_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
+
+  - name: cohere/command-r
+    display_name: Cohere Command R
+    description: Command R is a multilingual 35B parameter model with a context length of 128K that has been trained with conversational tool use capabilities.
+    creator_organization_name: Cohere
+    access: open
+    num_parameters: 35000000000
+    release_date: 2024-03-11
+    tags: [TEXT_MODEL_TAG, PARTIAL_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
+
+  - name: cohere/command-r-plus
+    display_name: Cohere Command R Plus
+    description: Command R+ is a multilingual 104B parameter model with a context length of 128K that has been trained with conversational tool use capabilities.
+    creator_organization_name: Cohere
+    access: open
+    num_parameters: 104000000000
+    release_date: 2024-04-04
+    tags: [TEXT_MODEL_TAG, PARTIAL_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
 
   # Craiyon
   - name: craiyon/dalle-mini

src/helm/config/tokenizer_configs.yaml

@@ -83,6 +83,22 @@ tokenizer_configs:
     end_of_text_token: ""
     prefix_token: ":"
 
+  - name: cohere/c4ai-command-r-v01
+    tokenizer_spec:
+      class_name: "helm.tokenizers.huggingface_tokenizer.HuggingFaceTokenizer"
+      args:
+        pretrained_model_name_or_path: CohereForAI/c4ai-command-r-v01
+    end_of_text_token: "<|END_OF_TURN_TOKEN|>"
+    prefix_token: "<BOS_TOKEN>"
+
+  - name: cohere/c4ai-command-r-plus
+    tokenizer_spec:
+      class_name: "helm.tokenizers.huggingface_tokenizer.HuggingFaceTokenizer"
+      args:
+        pretrained_model_name_or_path: CohereForAI/c4ai-command-r-plus
+    end_of_text_token: "<|END_OF_TURN_TOKEN|>"
+    prefix_token: "<BOS_TOKEN>"
+
   # Databricks
   - name: databricks/dbrx-instruct
     tokenizer_spec: