Add Phi 3.5 models (#3306)

stanford-crfm · Feb 1, 2025 · 228e0f1 · 228e0f1
1 parent f6a9856
commit 228e0f1
Show file tree

Hide file tree

Showing 3 changed files with 45 additions and 1 deletion.
diff --git a/src/helm/config/model_deployments.yaml b/src/helm/config/model_deployments.yaml
@@ -1129,7 +1129,27 @@ model_deployments:
       args:
         device_map: auto
         torch_dtype: auto
-
+
+  - name: huggingface/phi-3.5-mini-instruct
+    model_name: microsoft/phi-3.5-mini-instruct
+    tokenizer_name: microsoft/phi-3.5-mini-instruct
+    max_sequence_length: 131072
+    client_spec:
+      class_name: "helm.clients.huggingface_client.HuggingFaceClient"
+      args:
+        device_map: auto
+        torch_dtype: auto
+
+  - name: huggingface/phi-3.5-moe-instruct
+    model_name: microsoft/phi-3.5-moe-instruct
+    tokenizer_name: microsoft/phi-3.5-mini-instruct
+    max_sequence_length: 131072
+    client_spec:
+      class_name: "helm.clients.huggingface_client.HuggingFaceClient"
+      args:
+        device_map: auto
+        torch_dtype: auto
+
   ## Mistral AI
   - name: huggingface/bakLlava-v1-hf
     model_name: mistralai/bakLlava-v1-hf

diff --git a/src/helm/config/model_metadata.yaml b/src/helm/config/model_metadata.yaml
@@ -1899,6 +1899,24 @@ models:
     release_date: 2024-05-21
     tags: [TEXT_MODEL_TAG, LIMITED_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
 
+  - name: microsoft/phi-3.5-mini-instruct
+    display_name: Phi-3.5 Mini
+    description: Phi-3.5 Mini is a lightweight, state-of-the-art open model built upon datasets used for Phi-3 - synthetic data and filtered publicly available documents - with a focus on very high-quality, reasoning dense data. ([paper](https://arxiv.org/abs/2404.14219), [blog](https://techcommunity.microsoft.com/blog/azure-ai-services-blog/discover-the-new-multi-lingual-high-quality-phi-3-5-slms/4225280))
+    creator_organization_name: Microsoft
+    access: open
+    num_parameters: 3820000000
+    release_date: 2024-08-22
+    tags: [TEXT_MODEL_TAG, LIMITED_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
+
+  - name: microsoft/phi-3.5-moe-instruct
+    display_name: Phi-3.5 MoE
+    description: Phi-3.5 MoE is a lightweight, state-of-the-art open model built upon datasets used for Phi-3 - synthetic data and filtered publicly available documents - with a focus on very high-quality, reasoning dense data. ([paper](https://arxiv.org/abs/2404.14219), [blog](https://techcommunity.microsoft.com/blog/azure-ai-services-blog/discover-the-new-multi-lingual-high-quality-phi-3-5-slms/4225280))
+    creator_organization_name: Microsoft
+    access: open
+    num_parameters: 41900000000
+    release_date: 2024-08-22
+    tags: [TEXT_MODEL_TAG, LIMITED_FUNCTIONALITY_TEXT_MODEL_TAG, INSTRUCTION_FOLLOWING_MODEL_TAG]
+
   # KAIST AI
   - name: kaistai/prometheus-vision-13b-v1.0-hf
     display_name: LLaVA + Vicuna-v1.5 (13B)

diff --git a/src/helm/config/tokenizer_configs.yaml b/src/helm/config/tokenizer_configs.yaml
@@ -429,6 +429,12 @@ tokenizer_configs:
     end_of_text_token: "<|endoftext|>"
     prefix_token: "<s>"
 
+  - name: microsoft/phi-3.5-mini-instruct
+    tokenizer_spec:  
+      class_name: "helm.tokenizers.huggingface_tokenizer.HuggingFaceTokenizer"
+    end_of_text_token: "<|endoftext|>"
+    prefix_token: "<s>"
+
   # Mistralai
   - name: mistralai/Mistral-7B-v0.1
     tokenizer_spec: