Upload app.py with huggingface_hub

app.py

@@ -13,7 +13,8 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 BASE = "Qwen/Qwen2.5-3B-Instruct"
 LORA = "Curious-PM/lexwell-contract-irac-qwen2.5-3b-lora"
 
-SYSTEM_PROMPT = (
+BASE_PROMPT = "You are a helpful assistant."
+FT_PROMPT = (
     "You are an associate at Lexwell Advisors, a contract-review advisory "
     "firm for SMBs. Reply in Lexwell's house IRAC format with required top "
     "and bottom disclaimers."
@@ -33,22 +34,24 @@ def _decode(out, input_len):
     return tokenizer.decode(out[0][input_len:], skip_special_tokens=True)
 
 
+def _prepare_inputs(system_prompt, user_question):
+    msgs = [
+        {"role": "system", "content": system_prompt},
+        {"role": "user", "content": user_question.strip()},
+    ]
+    text = tokenizer.apply_chat_template(
+        msgs, tokenize=False, add_generation_prompt=True
+    )
+    return tokenizer(text, return_tensors="pt").to("cuda")
+
+
 @spaces.GPU(duration=120)
 def generate_both(question):
     if not question or not question.strip():
         placeholder = "_Type a contract question first._"
         return placeholder, placeholder
 
-    msgs = [
-        {"role": "system", "content": SYSTEM_PROMPT},
-        {"role": "user", "content": question.strip()},
-    ]
-    text = tokenizer.apply_chat_template(
-        msgs, tokenize=False, add_generation_prompt=True
-    )
-    inputs = tokenizer(text, return_tensors="pt").to("cuda")
     model.to("cuda")
-    input_len = inputs.input_ids.shape[1]
 
     gen_kwargs = dict(
         max_new_tokens=600,
@@ -56,16 +59,18 @@ def generate_both(question):
         pad_token_id=tokenizer.eos_token_id,
     )
 
-    # LEFT — base model (adapter disabled)
+    # LEFT — base model + generic "helpful assistant" prompt
+    base_inputs = _prepare_inputs(BASE_PROMPT, question)
     with model.disable_adapter():
         with torch.no_grad():
-            base_out = model.generate(**inputs, **gen_kwargs)
-    base_reply = _decode(base_out, input_len)
+            base_out = model.generate(**base_inputs, **gen_kwargs)
+    base_reply = _decode(base_out, base_inputs.input_ids.shape[1])
 
-    # RIGHT — fine-tuned (adapter on)
+    # RIGHT — fine-tuned + specific Lexwell prompt
+    ft_inputs = _prepare_inputs(FT_PROMPT, question)
     with torch.no_grad():
-        ft_out = model.generate(**inputs, **gen_kwargs)
-    ft_reply = _decode(ft_out, input_len)
+        ft_out = model.generate(**ft_inputs, **gen_kwargs)
+    ft_reply = _decode(ft_out, ft_inputs.input_ids.shape[1])
 
     return base_reply, ft_reply
 
@@ -371,7 +376,7 @@ with gr.Blocks(title="Junior Associate · Stay Curious", css=CSS, theme=gr.theme
             gr.HTML(
                 '<div class="lane-header">'
                 '  <div class="lane-title">Base Qwen 2.5-3B</div>'
-                '  <div class="lane-meta">no fine-tuning · prompt only</div>'
+                '  <div class="lane-meta">system: <code>&ldquo;You are a helpful assistant.&rdquo;</code></div>'
                 '</div>'
             )
             base_output = gr.Markdown(value=PLACEHOLDER, elem_classes="lane-body")
@@ -380,7 +385,7 @@ with gr.Blocks(title="Junior Associate · Stay Curious", css=CSS, theme=gr.theme
             gr.HTML(
                 '<div class="lane-header">'
                 '  <div class="lane-title">+ LoRA adapter (fine-tuned)</div>'
-                '  <div class="lane-meta">trained on 80 contract-review memos</div>'
+                '  <div class="lane-meta">system: <code>&ldquo;You are an associate at Lexwell Advisors&hellip;&rdquo;</code></div>'
                 '</div>'
             )
             ft_output = gr.Markdown(value=PLACEHOLDER, elem_classes="lane-body")