Fix mCORA, add Gemma preset

.gitignore

@@ -0,0 +1,3 @@
+*.pyc
+*.html
+*.json

app.py

@@ -20,6 +20,7 @@ from presets import (
     set_mmt_preset,
     set_towerinstruct_preset,
     set_zephyr_preset,
+    set_gemma_preset,
 )
 from style import custom_css
 from utils import get_formatted_attribute_context_results
@@ -115,7 +116,7 @@ def pecore(
     out = attribute_context_with_model(pecore_args, loaded_model)
     tuples = get_formatted_attribute_context_results(loaded_model, out.info, out)
     if not tuples:
-        msg = "Warning: No pairs were found by PECoRe. Try adjusting Results Selection parameters."
+        msg = f"Output: {out.output_current}\nWarning: No pairs were found by PECoRe. Try adjusting Results Selection parameters."
         tuples = [(msg, None)]
     return tuples, gr.Button(visible=True), gr.Button(visible=True)
 
@@ -217,7 +218,7 @@ with gr.Blocks(css=custom_css) as demo:
                     "Multilingual MT", variant="secondary"
                 )
                 gr.Markdown(
-                    "Present for multilingual MT models such as <a href='https://huggingface.co/facebook/nllb-200-distilled-600M' target='_blank'>NLLB</a> and <a href='https://huggingface.co/facebook/mbart-large-50-many-to-many-mmt' target='_blank'>mBART</a> using language tags."
+                    "Preset for multilingual MT models such as <a href='https://huggingface.co/facebook/nllb-200-distilled-600M' target='_blank'>NLLB</a> and <a href='https://huggingface.co/facebook/mbart-large-50-many-to-many-mmt' target='_blank'>mBART</a> using language tags."
                 )
             with gr.Column(scale=1):
                 chatml_template = gr.Button("Qwen ChatML", variant="secondary")
@@ -231,6 +232,14 @@ with gr.Blocks(css=custom_css) as demo:
                 gr.Markdown(
                     "Preset for models using the <a href='https://huggingface.co/Unbabel/TowerInstruct-7B-v0.1' target='_blank'>Unbabel TowerInstruct</a> conversational template.\nUses <code><|im_start|></code>, <code><|im_end|></code> special tokens."
                 )
+        with gr.Row(equal_height=True):
+            with gr.Column(scale=1):
+                gemma_template = gr.Button(
+                    "Gemma Chat Template", variant="secondary"
+                )
+                gr.Markdown(
+                    "Preset for <a href='https://huggingface.co/google/gemma-2b-it' target='_blank'>Gemma</a> instruction-tuned models."
+                )
         gr.Markdown("## ⚙️ PECoRe Parameters")
         with gr.Row(equal_height=True):
             with gr.Column():
@@ -512,6 +521,18 @@ with gr.Blocks(css=custom_css) as demo:
             input_template,
             contextless_input_current_text,
             decoder_input_output_separator,
+            special_tokens_to_keep,
+        ],
+    ).success(preload_model, inputs=load_model_args)
+
+    gemma_template.click(**reset_kwargs).then(
+        set_gemma_preset,
+        outputs=[
+            model_name_or_path,
+            input_template,
+            contextless_input_current_text,
+            decoder_input_output_separator,
+            special_tokens_to_keep,
         ],
     ).success(preload_model, inputs=load_model_args)
 

presets.py

@@ -55,4 +55,14 @@ def set_towerinstruct_preset():
         "<|im_start|>user\nSource: {current}\nContext: {context}\nTranslate the above text into French. Use the context to guide your answer.\nTarget:<|im_end|>\n<|im_start|>assistant\n",  # input_template
         "<|im_start|>user\nSource: {current}\nTranslate the above text into French.\nTarget:<|im_end|>\n<|im_start|>assistant\n",  # input_current_text_template
         "",  # decoder_input_output_separator
+        ["<|im_start|>", "<|im_end|>"],  # special_tokens_to_keep
+    )
+
+def set_gemma_preset():
+    return (
+        "google/gemma-2b-it", # model_name_or_path
+        "<start_of_turn>user\n{context}\n{current}<end_of_turn>\n<start_of_turn>model\n", # input_template
+        "<start_of_turn>user\n{current}<end_of_turn>\n<start_of_turn>model\n", # input_current_text_template
+        "", # decoder_input_output_separator
+        ["<start_of_turn>", "<end_of_turn>"], # special_tokens_to_keep
     )