Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
@@ -12,9 +12,9 @@ import time
|
|
12 |
login(token=os.environ["HF_TOKEN"])
|
13 |
|
14 |
# Structure hiérarchique des modèles
|
15 |
-
|
16 |
"meta-llama": {
|
17 |
-
"Llama-2": ["
|
18 |
"Llama-3": ["8B", "3.2-3B", "3.1-8B"]
|
19 |
},
|
20 |
"mistralai": {
|
@@ -22,7 +22,7 @@ models_hierarchy = {
|
|
22 |
"Mixtral": ["8x7B-v0.1"]
|
23 |
},
|
24 |
"google": {
|
25 |
-
"
|
26 |
},
|
27 |
"croissantllm": {
|
28 |
"CroissantLLM": ["Base"]
|
@@ -31,35 +31,35 @@ models_hierarchy = {
|
|
31 |
|
32 |
# Langues supportées par modèle
|
33 |
models_and_languages = {
|
34 |
-
"meta-llama/Llama-2-
|
35 |
-
"meta-llama/Llama-2-
|
36 |
-
"meta-llama/Llama-2-
|
37 |
-
"meta-llama/
|
38 |
"meta-llama/Llama-3.2-3B": ["en", "de", "fr", "it", "pt", "hi", "es", "th"],
|
39 |
"meta-llama/Llama-3.1-8B": ["en", "de", "fr", "it", "pt", "hi", "es", "th"],
|
40 |
"mistralai/Mistral-7B-v0.1": ["en"],
|
41 |
-
"mistralai/Mixtral-8x7B-v0.1": ["en", "fr", "it", "de", "es"],
|
42 |
"mistralai/Mistral-7B-v0.3": ["en"],
|
43 |
-
"
|
44 |
-
"google/
|
45 |
-
"google/
|
|
|
46 |
"croissantllm/CroissantLLMBase": ["en", "fr"]
|
47 |
}
|
48 |
|
49 |
# Paramètres recommandés pour chaque modèle
|
50 |
model_parameters = {
|
51 |
-
"meta-llama/Llama-2-
|
52 |
-
"meta-llama/Llama-2-
|
53 |
-
"meta-llama/Llama-2-
|
54 |
-
"meta-llama/
|
55 |
"meta-llama/Llama-3.2-3B": {"temperature": 0.75, "top_p": 0.9, "top_k": 50},
|
56 |
"meta-llama/Llama-3.1-8B": {"temperature": 0.75, "top_p": 0.9, "top_k": 50},
|
57 |
"mistralai/Mistral-7B-v0.1": {"temperature": 0.7, "top_p": 0.9, "top_k": 50},
|
58 |
-
"mistralai/Mixtral-8x7B-v0.1": {"temperature": 0.8, "top_p": 0.95, "top_k": 50},
|
59 |
"mistralai/Mistral-7B-v0.3": {"temperature": 0.7, "top_p": 0.9, "top_k": 50},
|
60 |
-
"
|
61 |
-
"google/
|
62 |
-
"google/
|
|
|
63 |
"croissantllm/CroissantLLMBase": {"temperature": 0.8, "top_p": 0.92, "top_k": 50}
|
64 |
}
|
65 |
|
@@ -69,24 +69,20 @@ tokenizer = None
|
|
69 |
selected_language = None
|
70 |
|
71 |
def update_model_choices(company):
|
72 |
-
return gr.Dropdown(choices=list(
|
73 |
|
74 |
def update_variation_choices(company, model_name):
|
75 |
-
return gr.Dropdown(choices=
|
76 |
|
77 |
def load_model(company, model_name, variation, progress=gr.Progress()):
|
78 |
global model, tokenizer
|
79 |
-
|
80 |
full_model_name = f"{company}/{model_name}-{variation}"
|
81 |
-
if full_model_name not in models_and_languages:
|
82 |
-
full_model_name = f"{company}/{model_name}{variation}"
|
83 |
|
84 |
try:
|
85 |
progress(0, desc="Chargement du tokenizer")
|
86 |
tokenizer = AutoTokenizer.from_pretrained(full_model_name)
|
87 |
progress(0.5, desc="Chargement du modèle")
|
88 |
|
89 |
-
# Configurations spécifiques par modèle
|
90 |
if "mixtral" in full_model_name.lower():
|
91 |
model = AutoModelForCausalLM.from_pretrained(
|
92 |
full_model_name,
|
@@ -106,9 +102,8 @@ def load_model(company, model_name, variation, progress=gr.Progress()):
|
|
106 |
|
107 |
progress(1.0, desc="Modèle chargé")
|
108 |
available_languages = models_and_languages[full_model_name]
|
109 |
-
|
110 |
-
# Mise à jour des sliders avec les valeurs recommandées
|
111 |
params = model_parameters[full_model_name]
|
|
|
112 |
return (
|
113 |
f"Modèle {full_model_name} chargé avec succès. Langues disponibles : {', '.join(available_languages)}",
|
114 |
gr.Dropdown(choices=available_languages, value=available_languages[0], visible=True, interactive=True),
|
@@ -119,15 +114,129 @@ def load_model(company, model_name, variation, progress=gr.Progress()):
|
|
119 |
except Exception as e:
|
120 |
return f"Erreur lors du chargement du modèle : {str(e)}", gr.Dropdown(visible=False), None, None, None
|
121 |
|
122 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
123 |
|
124 |
with gr.Blocks() as demo:
|
125 |
gr.Markdown("# LLM&BIAS")
|
126 |
|
127 |
with gr.Accordion("Sélection du modèle"):
|
128 |
-
company_dropdown = gr.Dropdown(choices=list(
|
129 |
-
model_dropdown = gr.Dropdown(label="Choisissez un modèle",
|
130 |
-
variation_dropdown = gr.Dropdown(label="Choisissez une variation",
|
131 |
load_button = gr.Button("Charger le modèle")
|
132 |
load_output = gr.Textbox(label="Statut du chargement")
|
133 |
language_dropdown = gr.Dropdown(label="Choisissez une langue", visible=False)
|
@@ -156,7 +265,7 @@ with gr.Blocks() as demo:
|
|
156 |
model_dropdown.change(update_variation_choices, inputs=[company_dropdown, model_dropdown], outputs=[variation_dropdown])
|
157 |
load_button.click(load_model,
|
158 |
inputs=[company_dropdown, model_dropdown, variation_dropdown],
|
159 |
-
outputs=[load_output, language_dropdown
|
160 |
language_dropdown.change(set_language, inputs=[language_dropdown], outputs=[language_output])
|
161 |
analyze_button.click(analyze_next_token,
|
162 |
inputs=[input_text, temperature, top_p, top_k],
|
|
|
12 |
login(token=os.environ["HF_TOKEN"])
|
13 |
|
14 |
# Structure hiérarchique des modèles
|
15 |
+
model_hierarchy = {
|
16 |
"meta-llama": {
|
17 |
+
"Llama-2": ["7B", "13B", "70B"],
|
18 |
"Llama-3": ["8B", "3.2-3B", "3.1-8B"]
|
19 |
},
|
20 |
"mistralai": {
|
|
|
22 |
"Mixtral": ["8x7B-v0.1"]
|
23 |
},
|
24 |
"google": {
|
25 |
+
"Gemma": ["2B", "9B", "27B"]
|
26 |
},
|
27 |
"croissantllm": {
|
28 |
"CroissantLLM": ["Base"]
|
|
|
31 |
|
32 |
# Langues supportées par modèle
|
33 |
models_and_languages = {
|
34 |
+
"meta-llama/Llama-2-7B": ["en"],
|
35 |
+
"meta-llama/Llama-2-13B": ["en"],
|
36 |
+
"meta-llama/Llama-2-70B": ["en"],
|
37 |
+
"meta-llama/Llama-3-8B": ["en"],
|
38 |
"meta-llama/Llama-3.2-3B": ["en", "de", "fr", "it", "pt", "hi", "es", "th"],
|
39 |
"meta-llama/Llama-3.1-8B": ["en", "de", "fr", "it", "pt", "hi", "es", "th"],
|
40 |
"mistralai/Mistral-7B-v0.1": ["en"],
|
|
|
41 |
"mistralai/Mistral-7B-v0.3": ["en"],
|
42 |
+
"mistralai/Mixtral-8x7B-v0.1": ["en", "fr", "it", "de", "es"],
|
43 |
+
"google/Gemma-2B": ["en"],
|
44 |
+
"google/Gemma-9B": ["en"],
|
45 |
+
"google/Gemma-27B": ["en"],
|
46 |
"croissantllm/CroissantLLMBase": ["en", "fr"]
|
47 |
}
|
48 |
|
49 |
# Paramètres recommandés pour chaque modèle
|
50 |
model_parameters = {
|
51 |
+
"meta-llama/Llama-2-7B": {"temperature": 0.8, "top_p": 0.9, "top_k": 40},
|
52 |
+
"meta-llama/Llama-2-13B": {"temperature": 0.8, "top_p": 0.9, "top_k": 40},
|
53 |
+
"meta-llama/Llama-2-70B": {"temperature": 0.8, "top_p": 0.9, "top_k": 40},
|
54 |
+
"meta-llama/Llama-3-8B": {"temperature": 0.75, "top_p": 0.9, "top_k": 50},
|
55 |
"meta-llama/Llama-3.2-3B": {"temperature": 0.75, "top_p": 0.9, "top_k": 50},
|
56 |
"meta-llama/Llama-3.1-8B": {"temperature": 0.75, "top_p": 0.9, "top_k": 50},
|
57 |
"mistralai/Mistral-7B-v0.1": {"temperature": 0.7, "top_p": 0.9, "top_k": 50},
|
|
|
58 |
"mistralai/Mistral-7B-v0.3": {"temperature": 0.7, "top_p": 0.9, "top_k": 50},
|
59 |
+
"mistralai/Mixtral-8x7B-v0.1": {"temperature": 0.8, "top_p": 0.95, "top_k": 50},
|
60 |
+
"google/Gemma-2B": {"temperature": 0.7, "top_p": 0.95, "top_k": 40},
|
61 |
+
"google/Gemma-9B": {"temperature": 0.7, "top_p": 0.95, "top_k": 40},
|
62 |
+
"google/Gemma-27B": {"temperature": 0.7, "top_p": 0.95, "top_k": 40},
|
63 |
"croissantllm/CroissantLLMBase": {"temperature": 0.8, "top_p": 0.92, "top_k": 50}
|
64 |
}
|
65 |
|
|
|
69 |
selected_language = None
|
70 |
|
71 |
def update_model_choices(company):
|
72 |
+
return gr.Dropdown(choices=list(model_hierarchy[company].keys()), value=None)
|
73 |
|
74 |
def update_variation_choices(company, model_name):
|
75 |
+
return gr.Dropdown(choices=model_hierarchy[company][model_name], value=None)
|
76 |
|
77 |
def load_model(company, model_name, variation, progress=gr.Progress()):
|
78 |
global model, tokenizer
|
|
|
79 |
full_model_name = f"{company}/{model_name}-{variation}"
|
|
|
|
|
80 |
|
81 |
try:
|
82 |
progress(0, desc="Chargement du tokenizer")
|
83 |
tokenizer = AutoTokenizer.from_pretrained(full_model_name)
|
84 |
progress(0.5, desc="Chargement du modèle")
|
85 |
|
|
|
86 |
if "mixtral" in full_model_name.lower():
|
87 |
model = AutoModelForCausalLM.from_pretrained(
|
88 |
full_model_name,
|
|
|
102 |
|
103 |
progress(1.0, desc="Modèle chargé")
|
104 |
available_languages = models_and_languages[full_model_name]
|
|
|
|
|
105 |
params = model_parameters[full_model_name]
|
106 |
+
|
107 |
return (
|
108 |
f"Modèle {full_model_name} chargé avec succès. Langues disponibles : {', '.join(available_languages)}",
|
109 |
gr.Dropdown(choices=available_languages, value=available_languages[0], visible=True, interactive=True),
|
|
|
114 |
except Exception as e:
|
115 |
return f"Erreur lors du chargement du modèle : {str(e)}", gr.Dropdown(visible=False), None, None, None
|
116 |
|
117 |
+
def set_language(lang):
|
118 |
+
global selected_language
|
119 |
+
selected_language = lang
|
120 |
+
return f"Langue sélectionnée : {lang}"
|
121 |
+
|
122 |
+
def ensure_token_display(token):
|
123 |
+
if token.isdigit() or (token.startswith('-') and token[1:].isdigit()):
|
124 |
+
return tokenizer.decode([int(token)])
|
125 |
+
return token
|
126 |
+
|
127 |
+
def analyze_next_token(input_text, temperature, top_p, top_k):
|
128 |
+
global model, tokenizer, selected_language
|
129 |
+
|
130 |
+
if model is None or tokenizer is None:
|
131 |
+
return "Veuillez d'abord charger un modèle.", None, None
|
132 |
+
|
133 |
+
inputs = tokenizer(input_text, return_tensors="pt", padding=True, truncation=True, max_length=512).to(model.device)
|
134 |
+
|
135 |
+
try:
|
136 |
+
with torch.no_grad():
|
137 |
+
outputs = model(**inputs)
|
138 |
+
|
139 |
+
last_token_logits = outputs.logits[0, -1, :]
|
140 |
+
probabilities = torch.nn.functional.softmax(last_token_logits, dim=-1)
|
141 |
+
|
142 |
+
top_k = 10
|
143 |
+
top_probs, top_indices = torch.topk(probabilities, top_k)
|
144 |
+
top_words = [ensure_token_display(tokenizer.decode([idx.item()])) for idx in top_indices]
|
145 |
+
prob_data = {word: prob.item() for word, prob in zip(top_words, top_probs)}
|
146 |
+
|
147 |
+
prob_text = "Prochains tokens les plus probables :\n\n"
|
148 |
+
for word, prob in prob_data.items():
|
149 |
+
prob_text += f"{word}: {prob:.2%}\n"
|
150 |
+
|
151 |
+
prob_plot = plot_probabilities(prob_data)
|
152 |
+
attention_plot = plot_attention(inputs["input_ids"][0].cpu(), last_token_logits.cpu())
|
153 |
+
|
154 |
+
return prob_text, attention_plot, prob_plot
|
155 |
+
except Exception as e:
|
156 |
+
return f"Erreur lors de l'analyse : {str(e)}", None, None
|
157 |
+
|
158 |
+
def generate_text(input_text, temperature, top_p, top_k):
|
159 |
+
global model, tokenizer, selected_language
|
160 |
+
|
161 |
+
if model is None or tokenizer is None:
|
162 |
+
return "Veuillez d'abord charger un modèle."
|
163 |
+
|
164 |
+
inputs = tokenizer(input_text, return_tensors="pt", padding=True, truncation=True, max_length=512).to(model.device)
|
165 |
+
|
166 |
+
try:
|
167 |
+
with torch.no_grad():
|
168 |
+
outputs = model.generate(
|
169 |
+
**inputs,
|
170 |
+
max_new_tokens=10,
|
171 |
+
temperature=temperature,
|
172 |
+
top_p=top_p,
|
173 |
+
top_k=top_k
|
174 |
+
)
|
175 |
+
|
176 |
+
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
177 |
+
return generated_text
|
178 |
+
except Exception as e:
|
179 |
+
return f"Erreur lors de la génération : {str(e)}"
|
180 |
+
|
181 |
+
def plot_probabilities(prob_data):
|
182 |
+
words = list(prob_data.keys())
|
183 |
+
probs = list(prob_data.values())
|
184 |
+
|
185 |
+
fig, ax = plt.subplots(figsize=(12, 6))
|
186 |
+
bars = ax.bar(range(len(words)), probs, color='lightgreen')
|
187 |
+
ax.set_title("Probabilités des tokens suivants les plus probables")
|
188 |
+
ax.set_xlabel("Tokens")
|
189 |
+
ax.set_ylabel("Probabilité")
|
190 |
+
|
191 |
+
ax.set_xticks(range(len(words)))
|
192 |
+
ax.set_xticklabels(words, rotation=45, ha='right')
|
193 |
+
|
194 |
+
for i, (bar, word) in enumerate(zip(bars, words)):
|
195 |
+
height = bar.get_height()
|
196 |
+
ax.text(i, height, f'{height:.2%}',
|
197 |
+
ha='center', va='bottom', rotation=0)
|
198 |
+
|
199 |
+
plt.tight_layout()
|
200 |
+
return fig
|
201 |
+
|
202 |
+
def plot_attention(input_ids, last_token_logits):
|
203 |
+
input_tokens = [ensure_token_display(tokenizer.decode([id])) for id in input_ids]
|
204 |
+
attention_scores = torch.nn.functional.softmax(last_token_logits, dim=-1)
|
205 |
+
top_k = min(len(input_tokens), 10)
|
206 |
+
top_attention_scores, _ = torch.topk(attention_scores, top_k)
|
207 |
+
|
208 |
+
fig, ax = plt.subplots(figsize=(14, 7))
|
209 |
+
sns.heatmap(top_attention_scores.unsqueeze(0).numpy(), annot=True, cmap="YlOrRd", cbar=True, ax=ax, fmt='.2%')
|
210 |
+
ax.set_xticklabels(input_tokens[-top_k:], rotation=45, ha="right", fontsize=10)
|
211 |
+
ax.set_yticklabels(["Attention"], rotation=0, fontsize=10)
|
212 |
+
ax.set_title("Scores d'attention pour les derniers tokens", fontsize=16)
|
213 |
+
|
214 |
+
cbar = ax.collections[0].colorbar
|
215 |
+
cbar.set_label("Score d'attention", fontsize=12)
|
216 |
+
cbar.ax.tick_params(labelsize=10)
|
217 |
+
|
218 |
+
plt.tight_layout()
|
219 |
+
return fig
|
220 |
+
|
221 |
+
def reset():
|
222 |
+
global model, tokenizer, selected_language
|
223 |
+
model = None
|
224 |
+
tokenizer = None
|
225 |
+
selected_language = None
|
226 |
+
return (
|
227 |
+
gr.Dropdown(choices=list(model_hierarchy.keys()), value=None),
|
228 |
+
gr.Dropdown(visible=False),
|
229 |
+
gr.Dropdown(visible=False),
|
230 |
+
"", 1.0, 1.0, 50, None, None, None, None, gr.Dropdown(visible=False), ""
|
231 |
+
)
|
232 |
|
233 |
with gr.Blocks() as demo:
|
234 |
gr.Markdown("# LLM&BIAS")
|
235 |
|
236 |
with gr.Accordion("Sélection du modèle"):
|
237 |
+
company_dropdown = gr.Dropdown(choices=list(model_hierarchy.keys()), label="Choisissez une société")
|
238 |
+
model_dropdown = gr.Dropdown(label="Choisissez un modèle", visible=False)
|
239 |
+
variation_dropdown = gr.Dropdown(label="Choisissez une variation", visible=False)
|
240 |
load_button = gr.Button("Charger le modèle")
|
241 |
load_output = gr.Textbox(label="Statut du chargement")
|
242 |
language_dropdown = gr.Dropdown(label="Choisissez une langue", visible=False)
|
|
|
265 |
model_dropdown.change(update_variation_choices, inputs=[company_dropdown, model_dropdown], outputs=[variation_dropdown])
|
266 |
load_button.click(load_model,
|
267 |
inputs=[company_dropdown, model_dropdown, variation_dropdown],
|
268 |
+
outputs=[load_output, language_dropdown])
|
269 |
language_dropdown.change(set_language, inputs=[language_dropdown], outputs=[language_output])
|
270 |
analyze_button.click(analyze_next_token,
|
271 |
inputs=[input_text, temperature, top_p, top_k],
|