loading model, control vectors, generating query

nlp/app/forms.py

@@ -17,11 +17,57 @@ class Meta:  # Ignoring CSRF security feature.
     )
 
     text_options = SelectField(label='Select Option:',
-        choices=[
-            ('option1', 'Option 1 Text'),
-            ('option2', 'Option 2 Text'),
-            ('option3', 'Option 3 Text'),
-        ],
+        choices= [
+    ('./nlp/data/vectors/Confident_Insecure.pt', 'Insecure - Confident'),
+    ('./nlp/data/vectors/Evasive_Direct.pt', 'Direct - Evasive'),
+    ('./nlp/data/vectors/Diplomatic_Blunt.pt', 'Blunt - Diplomatic'),
+    ('./nlp/data/vectors/Inquisitive_Disinterested.pt', 'Disinterested - Inquisitive'),
+    ('./nlp/data/vectors/Thoughtful_Impulsive.pt', 'Impulsive - Thoughtful'),
+    ('./nlp/data/vectors/Nonchalant_Concerned.pt', 'Concerned - Nonchalant'),
+    ('./nlp/data/vectors/Flippant_Serious.pt', 'Serious - Flippant'),
+    ('./nlp/data/vectors/Precise_Vague.pt', 'Vague - Precise'),
+    ('./nlp/data/vectors/Rambling_Concise.pt', 'Concise - Rambling'),
+    ('./nlp/data/vectors/Analytical_Intuitive.pt', 'Intuitive - Analytical'),
+    ('./nlp/data/vectors/Assertive_Passive.pt', 'Passive - Assertive'),
+    ('./nlp/data/vectors/Considerate_Thoughtless.pt', 'Thoughtless - Considerate'),
+    ('./nlp/data/vectors/Elusive_Clear.pt', 'Clear - Elusive'),
+    ('./nlp/data/vectors/Candid_Guarded.pt', 'Guarded - Candid'),
+    ('./nlp/data/vectors/Defensive_Open.pt', 'Open - Defensive'),
+    ('./nlp/data/vectors/Engaging_Detached.pt', 'Detached - Engaging'),
+    ('./nlp/data/vectors/Reserved_Outgoing.pt', 'Outgoing - Reserved'),
+    ('./nlp/data/vectors/Empathetic_Unsympathetic.pt', 'Unsympathetic - Empathetic'),
+    ('./nlp/data/vectors/Concise_Lengthy.pt', 'Lengthy - Concise'),
+    ('./nlp/data/vectors/Enthusiastic_Apathetic.pt', 'Apathetic - Enthusiastic'),
+    ('./nlp/data/vectors/Introspective_Extrospective.pt', 'Extrospective - Introspective'),
+    ('./nlp/data/vectors/Polite_Rude.pt', 'Rude - Polite'),
+    ('./nlp/data/vectors/Indecisive_Decisive.pt', 'Decisive - Indecisive'),
+    ('./nlp/data/vectors/Dismissive_Receptive.pt', 'Receptive - Dismissive'),
+    ('./nlp/data/vectors/Deliberate_Hasty.pt', 'Hasty - Deliberate'),
+    ('./nlp/data/vectors/Informative_Misleading.pt', 'Misleading - Informative'),
+    ('./nlp/data/vectors/Focused_Distracted.pt', 'Distracted - Focused'),
+    ('./nlp/data/vectors/Perplexed_Clear.pt', 'Clear - Perplexed'),
+    ('./nlp/data/vectors/Cooperative_Uncooperative.pt', 'Uncooperative - Cooperative'),
+    ('./nlp/data/vectors/Inattentive_Attentive.pt', 'Attentive - Inattentive'),
+    ('./nlp/data/vectors/Contemplative_Shallow.pt', 'Shallow - Contemplative'),
+    ('./nlp/data/vectors/Evocative_Uninspiring.pt', 'Uninspiring - Evocative'),
+    ('./nlp/data/vectors/Witty_Dull.pt', 'Dull - Witty'),
+    ('./nlp/data/vectors/Succinct_Rambling.pt', 'Rambling - Succinct'),
+    ('./nlp/data/vectors/Arrogant_Humble.pt', 'Humble - Arrogant'),
+    ('./nlp/data/vectors/Measured_Impulsive.pt', 'Impulsive - Measured'),
+    ('./nlp/data/vectors/Elaborate_Simple.pt', 'Simple - Elaborate'),
+    ('./nlp/data/vectors/Unresponsive_Responsive.pt', 'Responsive - Unresponsive'),
+    ('./nlp/data/vectors/Courteous_Rude.pt', 'Rude - Courteous'),
+    ('./nlp/data/vectors/Tentative_Definite.pt', 'Definite - Tentative'),
+    ('./nlp/data/vectors/Compelling_Unconvincing.pt', 'Unconvincing - Compelling'),
+    ('./nlp/data/vectors/Casual_Formal.pt', 'Formal - Casual'),
+    ('./nlp/data/vectors/Insightful_Superficial.pt', 'Superficial - Insightful'),
+    ('./nlp/data/vectors/Assertive_Passive_2.pt', 'Passive - Assertive'),
+    ('./nlp/data/vectors/Outgoing_Introverted.pt', 'Introverted - Outgoing'),
+    ('./nlp/data/vectors/Concise_Wordy.pt', 'Wordy - Concise'),
+    ('./nlp/data/vectors/Confident_Timid.pt', 'Timid - Confident'),
+    ('./nlp/data/vectors/Polite_Impolite.pt', 'Impolite - Polite'),
+    ('./nlp/data/vectors/Engaging_Aloof.pt', 'Aloof - Engaging')
+],
         validators=[DataRequired()],
         render_kw={'class': 'text-options-dropdown'}
     )

nlp/app/routes.py

@@ -29,6 +29,8 @@
 model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)
 model = model.to("cuda:0" if torch.cuda.is_available() else "cpu")
 
+model = ControlModel(model, list(range(-5, -18, -1)))
+user_tag, asst_tag = "[INST]", "[/INST]"
 
 
 
@@ -44,14 +46,44 @@ def index():
 	form = MyForm()
 	result = None
 	if form.validate_on_submit():
+		vector_path = form.text_options.data
+
+		control_vector = torch.load(vector_path)
+
 		input_field = form.input_field.data
+
+		input_query = user_tag + input_field + asst_tag
+
+		print("magnitude: %s" % form.magnitude.data)
+
+		mag = float(form.magnitude.data)
+
+		input_ids = tokenizer(input_query, return_tensors="pt").to(model.device)
+		settings = {
+			"pad_token_id": tokenizer.eos_token_id, # silence warning
+			"do_sample": False, # temperature=0
+			"max_new_tokens": 512,
+			"repetition_penalty": 1.1, # reduce control jank
+		}
+
+		print("==baseline")
+		model.reset()
+		default_output = str(tokenizer.decode(model.generate(**input_ids, **settings).squeeze()))
+
+
+		print("\n++control")
+		# add the control vector with a certain strength (try increasing or decreasing this!)
+		model.set_control(control_vector, mag)
+		control_output = str(tokenizer.decode(model.generate(**input_ids, **settings).squeeze()))
+
+		model.reset()
 		#X = vec.transform([input_field])
 		#pred = clf.predict(X)[0]
 		pred = "PRED"
 		#proba = clf.predict_proba(X)[0].max()
 		proba = 0.5
 		# flash(input_field)
 		return render_template('myform.html', title='', form=form, 
-								prediction=labels[pred], confidence='%.2f' % proba)
+								prediction=control_output, confidence='%.2f' % proba)
 		#return redirect('/index')
 	return render_template('myform.html', title='', form=form, prediction=None, confidence=None)