Spaces:
Runtime error
Runtime error
frankaging
commited on
Commit
·
e39562b
1
Parent(s):
1baa5c3
o1 impl
Browse files
app.py
CHANGED
|
@@ -12,22 +12,8 @@ HF_TOKEN = os.environ.get("HF_TOKEN")
|
|
| 12 |
login(token=HF_TOKEN)
|
| 13 |
|
| 14 |
MAX_MAX_NEW_TOKENS = 2048
|
| 15 |
-
DEFAULT_MAX_NEW_TOKENS =
|
| 16 |
-
MAX_INPUT_TOKEN_LENGTH =
|
| 17 |
-
|
| 18 |
-
DESCRIPTION = """\
|
| 19 |
-
# Model Steering with Supervised Dictionary Learning (SDL)
|
| 20 |
-
|
| 21 |
-
### What's Model Steering with SDL?
|
| 22 |
-
This is a demo of model steering with AxBench-ReFT-r1-16K, ...
|
| 23 |
-
"""
|
| 24 |
-
|
| 25 |
-
LICENSE = """
|
| 26 |
-
<p/>
|
| 27 |
-
|
| 28 |
-
---
|
| 29 |
-
Please refer to the specific licensing and use policy of the underlying model.
|
| 30 |
-
"""
|
| 31 |
|
| 32 |
def load_jsonl(jsonl_path):
|
| 33 |
jsonl_data = []
|
|
@@ -38,41 +24,41 @@ def load_jsonl(jsonl_path):
|
|
| 38 |
return jsonl_data
|
| 39 |
|
| 40 |
class Steer(pv.SourcelessIntervention):
|
| 41 |
-
"""Steer model via activation addition"""
|
| 42 |
def __init__(self, **kwargs):
|
| 43 |
super().__init__(**kwargs, keep_last_dim=True)
|
| 44 |
self.proj = torch.nn.Linear(self.embed_dim, kwargs["latent_dim"], bias=False)
|
| 45 |
|
| 46 |
def forward(self, base, source=None, subspaces=None):
|
| 47 |
-
# subspaces is a list of dicts:
|
| 48 |
-
# each has {"idx": int, "internal_mag": float, "text": str, ...}
|
| 49 |
steer_vec = base
|
| 50 |
if subspaces is not None:
|
| 51 |
for sp in subspaces:
|
| 52 |
idx = sp["idx"]
|
| 53 |
-
mag = sp["internal_mag"] #
|
| 54 |
steering_vec = mag * self.proj.weight[idx].unsqueeze(dim=0)
|
| 55 |
steer_vec = steer_vec + steering_vec
|
| 56 |
return steer_vec
|
| 57 |
|
| 58 |
-
|
| 59 |
-
# ------------------------------------------
|
| 60 |
-
# Load the Model & Dictionary if GPU exists
|
| 61 |
-
# ------------------------------------------
|
| 62 |
if not torch.cuda.is_available():
|
| 63 |
-
|
| 64 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
if torch.cuda.is_available():
|
| 66 |
-
model_id = "google/gemma-2-2b-it"
|
| 67 |
model = AutoModelForCausalLM.from_pretrained(
|
| 68 |
model_id, device_map="cuda", torch_dtype=torch.bfloat16
|
| 69 |
)
|
| 70 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 71 |
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
|
|
|
| 76 |
|
| 77 |
concept_list = [item["concept"] for item in md]
|
| 78 |
concept_id_map = {item["concept"]: item["concept_id"] for item in md}
|
|
@@ -88,12 +74,8 @@ if torch.cuda.is_available():
|
|
| 88 |
model=model,
|
| 89 |
)
|
| 90 |
|
| 91 |
-
terminators = [tokenizer.eos_token_id]
|
| 92 |
-
|
| 93 |
|
| 94 |
-
# --------------------------------------------
|
| 95 |
-
# Main generation function: keep last 3 turns
|
| 96 |
-
# --------------------------------------------
|
| 97 |
@spaces.GPU
|
| 98 |
def generate(
|
| 99 |
message: str,
|
|
@@ -101,37 +83,28 @@ def generate(
|
|
| 101 |
max_new_tokens: int,
|
| 102 |
subspaces_list: list[dict],
|
| 103 |
) -> Iterator[str]:
|
| 104 |
-
|
| 105 |
-
# Restrict to the last 3 turns only
|
| 106 |
start_idx = max(0, len(chat_history) - 3)
|
| 107 |
recent_history = chat_history[start_idx:]
|
| 108 |
|
| 109 |
-
#
|
| 110 |
messages = []
|
| 111 |
for user_msg, model_msg in recent_history:
|
| 112 |
messages.append({"role": "user", "content": user_msg})
|
| 113 |
messages.append({"role": "model", "content": model_msg})
|
| 114 |
-
|
| 115 |
-
# Add the new user message
|
| 116 |
messages.append({"role": "user", "content": message})
|
| 117 |
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
prompt_dict = tokenizer.apply_chat_template(
|
| 121 |
-
messages, tokenize=True, add_generation_prompt=True
|
| 122 |
-
)
|
| 123 |
-
input_ids = torch.tensor([prompt_dict["input_ids"]]).cuda()
|
| 124 |
-
attention_mask = torch.tensor([prompt_dict["attention_mask"]]).cuda()
|
| 125 |
|
| 126 |
-
#
|
| 127 |
if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
|
| 128 |
input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
|
| 129 |
-
|
| 130 |
-
yield "\n[Warning: Truncated conversation exceeds max allowed input tokens]\n"
|
| 131 |
|
| 132 |
streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
|
| 133 |
generate_kwargs = {
|
| 134 |
-
"base": {"input_ids": input_ids
|
| 135 |
"unit_locations": None,
|
| 136 |
"max_new_tokens": max_new_tokens,
|
| 137 |
"intervene_on_prompt": True,
|
|
@@ -150,29 +123,20 @@ def generate(
|
|
| 150 |
partial_text.append(token_str)
|
| 151 |
yield "".join(partial_text)
|
| 152 |
|
| 153 |
-
|
| 154 |
-
# ----------------
|
| 155 |
-
# UI Callbacks
|
| 156 |
-
# ----------------
|
| 157 |
def filter_concepts(search_text: str):
|
| 158 |
-
"""Return the first 500 concepts that match (case-insensitive)."""
|
| 159 |
if not search_text.strip():
|
| 160 |
return concept_list[:500]
|
| 161 |
filtered = [c for c in concept_list if search_text.lower() in c.lower()]
|
| 162 |
return filtered[:500]
|
| 163 |
|
| 164 |
def add_concept_to_list(selected_concept, user_slider_val, current_list):
|
| 165 |
-
"""
|
| 166 |
-
user_slider_val is from [-5..5]. We multiply by 50 internally to get the real magnitude.
|
| 167 |
-
"""
|
| 168 |
if not selected_concept:
|
| 169 |
return current_list, _build_table_data(current_list), gr.update(choices=_build_remove_choices(current_list))
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
internal_mag = user_slider_val * 50 # scale by 50
|
| 173 |
new_entry = {
|
| 174 |
"text": selected_concept,
|
| 175 |
-
"idx":
|
| 176 |
"display_mag": user_slider_val,
|
| 177 |
"internal_mag": internal_mag,
|
| 178 |
}
|
|
@@ -183,14 +147,10 @@ def add_concept_to_list(selected_concept, user_slider_val, current_list):
|
|
| 183 |
gr.update(choices=_build_remove_choices(updated_list))
|
| 184 |
)
|
| 185 |
|
| 186 |
-
def remove_concept_from_list(
|
| 187 |
-
|
| 188 |
-
Remove the chosen concept name from the list.
|
| 189 |
-
"""
|
| 190 |
-
if not concept_to_remove:
|
| 191 |
return current_list, _build_table_data(current_list), gr.update(choices=_build_remove_choices(current_list))
|
| 192 |
-
|
| 193 |
-
updated_list = [x for x in current_list if x["text"] != concept_to_remove]
|
| 194 |
return (
|
| 195 |
updated_list,
|
| 196 |
_build_table_data(updated_list),
|
|
@@ -198,115 +158,104 @@ def remove_concept_from_list(concept_to_remove, current_list):
|
|
| 198 |
)
|
| 199 |
|
| 200 |
def _build_table_data(subspaces):
|
| 201 |
-
"""Return [[concept_name, scaled_mag], ...] for display."""
|
| 202 |
return [[x["text"], x["display_mag"]] for x in subspaces]
|
| 203 |
|
| 204 |
def _build_remove_choices(subspaces):
|
| 205 |
-
"""Return concept names for the remove dropdown."""
|
| 206 |
return [x["text"] for x in subspaces]
|
| 207 |
|
| 208 |
def update_dropdown_choices(search_text):
|
| 209 |
filtered = filter_concepts(search_text)
|
| 210 |
return gr.update(choices=filtered)
|
| 211 |
|
| 212 |
-
# --------------------------------------------------------------------
|
| 213 |
-
# Build the Interface
|
| 214 |
-
# --------------------------------------------------------------------
|
| 215 |
with gr.Blocks(css="style.css") as demo:
|
| 216 |
-
|
| 217 |
-
gr.
|
| 218 |
|
| 219 |
-
#
|
| 220 |
default_subspaces = []
|
| 221 |
-
if
|
| 222 |
default_concept = random.choice(concept_list)
|
| 223 |
default_subspaces = [{
|
| 224 |
"text": default_concept,
|
| 225 |
"idx": concept_id_map[default_concept],
|
| 226 |
-
"display_mag": 3,
|
| 227 |
-
"internal_mag": 150.0,
|
| 228 |
}]
|
| 229 |
|
| 230 |
selected_subspaces = gr.State(default_subspaces)
|
| 231 |
-
|
| 232 |
with gr.Row():
|
| 233 |
-
|
| 234 |
-
|
| 235 |
chat_interface = gr.ChatInterface(
|
| 236 |
fn=generate,
|
| 237 |
-
additional_inputs=[
|
| 238 |
-
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
maximum=MAX_MAX_NEW_TOKENS,
|
| 242 |
-
step=1,
|
| 243 |
-
value=DEFAULT_MAX_NEW_TOKENS,
|
| 244 |
-
),
|
| 245 |
-
selected_subspaces
|
| 246 |
-
],
|
| 247 |
-
title="Model Steering with ReFT-r1 (16K concepts)",
|
| 248 |
-
type="messages", # <--- uses openai-style 'role' and 'content'
|
| 249 |
)
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
search_box = gr.Textbox(
|
| 254 |
label="Search concepts",
|
| 255 |
-
placeholder="
|
| 256 |
)
|
| 257 |
concept_dropdown = gr.Dropdown(
|
| 258 |
label="Filtered Concepts",
|
| 259 |
-
choices=[]
|
| 260 |
-
multiselect=False
|
| 261 |
)
|
| 262 |
concept_magnitude = gr.Slider(
|
| 263 |
-
label="Scaled
|
| 264 |
minimum=-5,
|
| 265 |
maximum=5,
|
| 266 |
step=1,
|
| 267 |
value=3
|
| 268 |
)
|
| 269 |
add_button = gr.Button("Add Concept")
|
| 270 |
-
|
| 271 |
-
# Show the table of active subspaces
|
| 272 |
active_subspaces_table = gr.Dataframe(
|
| 273 |
-
headers=["Concept", "
|
| 274 |
datatype=["str", "number"],
|
| 275 |
value=_build_table_data(default_subspaces),
|
| 276 |
interactive=False,
|
| 277 |
-
label="Active Concept Subspaces"
|
|
|
|
| 278 |
)
|
| 279 |
-
|
| 280 |
-
|
| 281 |
-
|
| 282 |
-
|
| 283 |
-
|
| 284 |
-
|
| 285 |
-
|
| 286 |
-
|
| 287 |
-
|
| 288 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 289 |
|
| 290 |
# Wire up events
|
| 291 |
-
|
| 292 |
-
search_box.change(
|
| 293 |
-
fn=update_dropdown_choices,
|
| 294 |
-
inputs=[search_box],
|
| 295 |
-
outputs=[concept_dropdown]
|
| 296 |
-
)
|
| 297 |
-
|
| 298 |
-
# Add concept
|
| 299 |
add_button.click(
|
| 300 |
-
|
| 301 |
-
|
| 302 |
-
|
| 303 |
)
|
| 304 |
-
|
| 305 |
-
# Remove a concept
|
| 306 |
remove_button.click(
|
| 307 |
-
|
| 308 |
-
|
| 309 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 310 |
)
|
| 311 |
|
| 312 |
-
demo.
|
|
|
|
| 12 |
login(token=HF_TOKEN)
|
| 13 |
|
| 14 |
MAX_MAX_NEW_TOKENS = 2048
|
| 15 |
+
DEFAULT_MAX_NEW_TOKENS = 512 # smaller default to save memory
|
| 16 |
+
MAX_INPUT_TOKEN_LENGTH = 4096
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
|
| 18 |
def load_jsonl(jsonl_path):
|
| 19 |
jsonl_data = []
|
|
|
|
| 24 |
return jsonl_data
|
| 25 |
|
| 26 |
class Steer(pv.SourcelessIntervention):
|
|
|
|
| 27 |
def __init__(self, **kwargs):
|
| 28 |
super().__init__(**kwargs, keep_last_dim=True)
|
| 29 |
self.proj = torch.nn.Linear(self.embed_dim, kwargs["latent_dim"], bias=False)
|
| 30 |
|
| 31 |
def forward(self, base, source=None, subspaces=None):
|
|
|
|
|
|
|
| 32 |
steer_vec = base
|
| 33 |
if subspaces is not None:
|
| 34 |
for sp in subspaces:
|
| 35 |
idx = sp["idx"]
|
| 36 |
+
mag = sp["internal_mag"] # scaled by 50
|
| 37 |
steering_vec = mag * self.proj.weight[idx].unsqueeze(dim=0)
|
| 38 |
steer_vec = steer_vec + steering_vec
|
| 39 |
return steer_vec
|
| 40 |
|
| 41 |
+
# Check GPU
|
|
|
|
|
|
|
|
|
|
| 42 |
if not torch.cuda.is_available():
|
| 43 |
+
print("Warning: Running on CPU, may be slow.")
|
| 44 |
+
|
| 45 |
+
# Load model & dictionary
|
| 46 |
+
model_id = "google/gemma-2-2b-it"
|
| 47 |
+
pv_model = None
|
| 48 |
+
tokenizer = None
|
| 49 |
+
concept_list = []
|
| 50 |
+
concept_id_map = {}
|
| 51 |
if torch.cuda.is_available():
|
|
|
|
| 52 |
model = AutoModelForCausalLM.from_pretrained(
|
| 53 |
model_id, device_map="cuda", torch_dtype=torch.bfloat16
|
| 54 |
)
|
| 55 |
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 56 |
|
| 57 |
+
# Download dictionary
|
| 58 |
+
weight_path = hf_hub_download(repo_id="pyvene/gemma-reft-2b-it-res", filename="l20/weight.pt")
|
| 59 |
+
meta_path = hf_hub_download(repo_id="pyvene/gemma-reft-2b-it-res", filename="l20/metadata.jsonl")
|
| 60 |
+
params = torch.load(weight_path).cuda()
|
| 61 |
+
md = load_jsonl(meta_path)
|
| 62 |
|
| 63 |
concept_list = [item["concept"] for item in md]
|
| 64 |
concept_id_map = {item["concept"]: item["concept_id"] for item in md}
|
|
|
|
| 74 |
model=model,
|
| 75 |
)
|
| 76 |
|
| 77 |
+
terminators = [tokenizer.eos_token_id] if tokenizer else []
|
|
|
|
| 78 |
|
|
|
|
|
|
|
|
|
|
| 79 |
@spaces.GPU
|
| 80 |
def generate(
|
| 81 |
message: str,
|
|
|
|
| 83 |
max_new_tokens: int,
|
| 84 |
subspaces_list: list[dict],
|
| 85 |
) -> Iterator[str]:
|
| 86 |
+
# limit to last 3 turns
|
|
|
|
| 87 |
start_idx = max(0, len(chat_history) - 3)
|
| 88 |
recent_history = chat_history[start_idx:]
|
| 89 |
|
| 90 |
+
# build list of messages
|
| 91 |
messages = []
|
| 92 |
for user_msg, model_msg in recent_history:
|
| 93 |
messages.append({"role": "user", "content": user_msg})
|
| 94 |
messages.append({"role": "model", "content": model_msg})
|
|
|
|
|
|
|
| 95 |
messages.append({"role": "user", "content": message})
|
| 96 |
|
| 97 |
+
input_ids = torch.tensor([tokenizer.apply_chat_template(
|
| 98 |
+
messages, tokenize=True, add_generation_prompt=True)]).cuda()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 99 |
|
| 100 |
+
# trim if needed
|
| 101 |
if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
|
| 102 |
input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
|
| 103 |
+
yield "[Truncated prior text]\n"
|
|
|
|
| 104 |
|
| 105 |
streamer = TextIteratorStreamer(tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True)
|
| 106 |
generate_kwargs = {
|
| 107 |
+
"base": {"input_ids": input_ids},
|
| 108 |
"unit_locations": None,
|
| 109 |
"max_new_tokens": max_new_tokens,
|
| 110 |
"intervene_on_prompt": True,
|
|
|
|
| 123 |
partial_text.append(token_str)
|
| 124 |
yield "".join(partial_text)
|
| 125 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 126 |
def filter_concepts(search_text: str):
|
|
|
|
| 127 |
if not search_text.strip():
|
| 128 |
return concept_list[:500]
|
| 129 |
filtered = [c for c in concept_list if search_text.lower() in c.lower()]
|
| 130 |
return filtered[:500]
|
| 131 |
|
| 132 |
def add_concept_to_list(selected_concept, user_slider_val, current_list):
|
|
|
|
|
|
|
|
|
|
| 133 |
if not selected_concept:
|
| 134 |
return current_list, _build_table_data(current_list), gr.update(choices=_build_remove_choices(current_list))
|
| 135 |
+
idx = concept_id_map[selected_concept]
|
| 136 |
+
internal_mag = user_slider_val * 50
|
|
|
|
| 137 |
new_entry = {
|
| 138 |
"text": selected_concept,
|
| 139 |
+
"idx": idx,
|
| 140 |
"display_mag": user_slider_val,
|
| 141 |
"internal_mag": internal_mag,
|
| 142 |
}
|
|
|
|
| 147 |
gr.update(choices=_build_remove_choices(updated_list))
|
| 148 |
)
|
| 149 |
|
| 150 |
+
def remove_concept_from_list(selected_text, current_list):
|
| 151 |
+
if not selected_text:
|
|
|
|
|
|
|
|
|
|
| 152 |
return current_list, _build_table_data(current_list), gr.update(choices=_build_remove_choices(current_list))
|
| 153 |
+
updated_list = [x for x in current_list if x["text"] != selected_text]
|
|
|
|
| 154 |
return (
|
| 155 |
updated_list,
|
| 156 |
_build_table_data(updated_list),
|
|
|
|
| 158 |
)
|
| 159 |
|
| 160 |
def _build_table_data(subspaces):
|
|
|
|
| 161 |
return [[x["text"], x["display_mag"]] for x in subspaces]
|
| 162 |
|
| 163 |
def _build_remove_choices(subspaces):
|
|
|
|
| 164 |
return [x["text"] for x in subspaces]
|
| 165 |
|
| 166 |
def update_dropdown_choices(search_text):
|
| 167 |
filtered = filter_concepts(search_text)
|
| 168 |
return gr.update(choices=filtered)
|
| 169 |
|
|
|
|
|
|
|
|
|
|
| 170 |
with gr.Blocks(css="style.css") as demo:
|
| 171 |
+
# A short title only
|
| 172 |
+
gr.Markdown("## Model Steering with ReFT-r1 (16K concepts)")
|
| 173 |
|
| 174 |
+
# Pre-populate with a random concept if available
|
| 175 |
default_subspaces = []
|
| 176 |
+
if pv_model and concept_list:
|
| 177 |
default_concept = random.choice(concept_list)
|
| 178 |
default_subspaces = [{
|
| 179 |
"text": default_concept,
|
| 180 |
"idx": concept_id_map[default_concept],
|
| 181 |
+
"display_mag": 3,
|
| 182 |
+
"internal_mag": 150.0,
|
| 183 |
}]
|
| 184 |
|
| 185 |
selected_subspaces = gr.State(default_subspaces)
|
|
|
|
| 186 |
with gr.Row():
|
| 187 |
+
# Left side: bigger chat area
|
| 188 |
+
with gr.Column(scale=7):
|
| 189 |
chat_interface = gr.ChatInterface(
|
| 190 |
fn=generate,
|
| 191 |
+
additional_inputs=[], # we'll put the max tokens slider below
|
| 192 |
+
title="",
|
| 193 |
+
type="messages",
|
| 194 |
+
height=550 # a bit taller to show more conversation
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 195 |
)
|
| 196 |
+
# Right side: concept management
|
| 197 |
+
with gr.Column(scale=3):
|
| 198 |
+
gr.Markdown("### Steering Concepts")
|
| 199 |
search_box = gr.Textbox(
|
| 200 |
label="Search concepts",
|
| 201 |
+
placeholder="e.g. 'time travel'"
|
| 202 |
)
|
| 203 |
concept_dropdown = gr.Dropdown(
|
| 204 |
label="Filtered Concepts",
|
| 205 |
+
choices=[]
|
|
|
|
| 206 |
)
|
| 207 |
concept_magnitude = gr.Slider(
|
| 208 |
+
label="Scaled Factor",
|
| 209 |
minimum=-5,
|
| 210 |
maximum=5,
|
| 211 |
step=1,
|
| 212 |
value=3
|
| 213 |
)
|
| 214 |
add_button = gr.Button("Add Concept")
|
|
|
|
|
|
|
| 215 |
active_subspaces_table = gr.Dataframe(
|
| 216 |
+
headers=["Concept", "Mag (scaled)"],
|
| 217 |
datatype=["str", "number"],
|
| 218 |
value=_build_table_data(default_subspaces),
|
| 219 |
interactive=False,
|
| 220 |
+
label="Active Concept Subspaces",
|
| 221 |
+
height=170 # give it a bit more room
|
| 222 |
)
|
| 223 |
+
# Row with the remove dropdown + button
|
| 224 |
+
with gr.Row():
|
| 225 |
+
remove_dropdown = gr.Dropdown(
|
| 226 |
+
label="Remove concept",
|
| 227 |
+
choices=_build_remove_choices(default_subspaces),
|
| 228 |
+
multiselect=False
|
| 229 |
+
)
|
| 230 |
+
remove_button = gr.Button("Remove", variant="secondary")
|
| 231 |
+
|
| 232 |
+
# Place the max tokens slider at bottom, smaller
|
| 233 |
+
with gr.Row():
|
| 234 |
+
gr.Markdown("**Max New Tokens**", elem_classes=["small-label"])
|
| 235 |
+
max_token_slider = gr.Slider(
|
| 236 |
+
minimum=1, maximum=MAX_MAX_NEW_TOKENS, step=1,
|
| 237 |
+
value=DEFAULT_MAX_NEW_TOKENS,
|
| 238 |
+
label="", # hide the big label
|
| 239 |
+
container=False,
|
| 240 |
+
style={"width": "30%"} # narrower
|
| 241 |
+
)
|
| 242 |
|
| 243 |
# Wire up events
|
| 244 |
+
search_box.change(update_dropdown_choices, [search_box], [concept_dropdown])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 245 |
add_button.click(
|
| 246 |
+
add_concept_to_list,
|
| 247 |
+
[concept_dropdown, concept_magnitude, selected_subspaces],
|
| 248 |
+
[selected_subspaces, active_subspaces_table, remove_dropdown]
|
| 249 |
)
|
|
|
|
|
|
|
| 250 |
remove_button.click(
|
| 251 |
+
remove_concept_from_list,
|
| 252 |
+
[remove_dropdown, selected_subspaces],
|
| 253 |
+
[selected_subspaces, active_subspaces_table, remove_dropdown]
|
| 254 |
+
)
|
| 255 |
+
|
| 256 |
+
# Link the slider back to chat generation
|
| 257 |
+
chat_interface.configure(
|
| 258 |
+
extra_inputs=[max_token_slider, selected_subspaces]
|
| 259 |
)
|
| 260 |
|
| 261 |
+
demo.launch()
|