Add new Ministral and Mistral Small models (#3110)

stanford-crfm · Oct 29, 2024 · 0a7273a · 0a7273a
1 parent 9376c1c
commit 0a7273a
Show file tree

Hide file tree

Showing 3 changed files with 49 additions and 6 deletions.
diff --git a/src/helm/config/model_deployments.yaml b/src/helm/config/model_deployments.yaml
@@ -1498,14 +1498,20 @@ model_deployments:
         precision: bf16-true
 
   # Mistral AI
-  - name: mistralai/mistral-tiny
-    model_name: mistralai/mistral-7b-v0.1
-    tokenizer_name: mistralai/Mistral-7B-v0.1
-    max_sequence_length: 32000
+
+  - name: mistralai/ministral-3b-2410
+    model_name: mistralai/ministral-3b-2410
+    tokenizer_name: mistralai/Ministral-8B-Instruct-2410
+    max_sequence_length: 128000
+    client_spec:
+      class_name: "helm.clients.mistral_client.MistralAIClient"
+
+  - name: mistralai/ministral-8b-2410
+    model_name: mistralai/ministral-8b-2410
+    tokenizer_name: mistralai/Ministral-8B-Instruct-2410
+    max_sequence_length: 128000
     client_spec:
       class_name: "helm.clients.mistral_client.MistralAIClient"
-      args:
-        mistral_model: "mistral-tiny"
 
   - name: mistralai/mistral-small-2402
     model_name: mistralai/mistral-small-2402
@@ -1514,6 +1520,13 @@ model_deployments:
     client_spec:
       class_name: "helm.clients.mistral_client.MistralAIClient"
 
+  - name: mistralai/mistral-small-2409
+    model_name: mistralai/mistral-small-2409
+    tokenizer_name: mistralai/Mistral-7B-v0.1
+    max_sequence_length: 32000
+    client_spec:
+      class_name: "helm.clients.mistral_client.MistralAIClient"
+
   - name: mistralai/mistral-medium-2312
     model_name: mistralai/mistral-medium-2312
     tokenizer_name: mistralai/Mistral-7B-v0.1

diff --git a/src/helm/config/model_metadata.yaml b/src/helm/config/model_metadata.yaml
@@ -1988,6 +1988,22 @@ models:
     release_date: 2023-10-16
     tags: [VISION_LANGUAGE_MODEL_TAG, LLAVA_MODEL_TAG, LIMITED_FUNCTIONALITY_VLM_TAG]
 
+  - name: mistralai/ministral-3b-2410
+    display_name: Ministral 3B (2402)
+    description: Ministral 3B (2402) is a model for on-device computing and at-the-edge use cases ([blog](https://mistral.ai/news/ministraux/)).
+    creator_organization_name: Mistral AI
+    access: limited
+    release_date: 2024-10-16
+    tags: [TEXT_MODEL_TAG, LIMITED_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
+
+  - name: mistralai/ministral-8b-2410
+    display_name: Ministral 8B (2402)
+    description: Ministral 8B (2402) is a model for on-device computing and at-the-edge use cases a special interleaved sliding-window attention pattern for faster and memory-efficient inference ([blog](https://mistral.ai/news/ministraux/)).
+    creator_organization_name: Mistral AI
+    access: open
+    release_date: 2024-10-16
+    tags: [TEXT_MODEL_TAG, LIMITED_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
+
   - name: mistralai/mistral-small-2402
     display_name: Mistral Small (2402)
     description: Mistral Small is a multilingual model with a 32K tokens context window and function-calling capabilities. ([blog](https://mistral.ai/news/mistral-large/))
@@ -1996,6 +2012,14 @@ models:
     release_date: 2023-02-26
     tags: [TEXT_MODEL_TAG, LIMITED_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
 
+  - name: mistralai/mistral-small-2409
+    display_name: Mistral Small (2409)
+    description: Mistral Small is a multilingual model with a 32K tokens context window and function-calling capabilities. ([blog](https://mistral.ai/news/mistral-large/))
+    creator_organization_name: Mistral AI
+    access: limited
+    release_date: 2024-09-18
+    tags: [TEXT_MODEL_TAG, LIMITED_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
+
   - name: mistralai/mistral-medium-2312
     display_name: Mistral Medium (2312)
     description: Mistral is a transformer model that uses Grouped-Query Attention (GQA) and Sliding-Window Attention (SWA).

diff --git a/src/helm/config/tokenizer_configs.yaml b/src/helm/config/tokenizer_configs.yaml
@@ -444,6 +444,12 @@ tokenizer_configs:
     end_of_text_token: "</s>"
     prefix_token: "<s>"
 
+  - name: mistralai/Ministral-8B-Instruct-2410
+    tokenizer_spec:
+      class_name: "helm.tokenizers.huggingface_tokenizer.HuggingFaceTokenizer"
+    end_of_text_token: "</s>"
+    prefix_token: "<s>"
+
   # Neurips
   - name: neurips/local
     tokenizer_spec: