Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,51 +1,130 @@
|
|
| 1 |
import time
|
| 2 |
-
from typing import List, Tuple,
|
| 3 |
|
| 4 |
import gradio as gr
|
| 5 |
-
import numpy as np
|
| 6 |
-
from sklearn.metrics.pairwise import cosine_similarity
|
| 7 |
-
import torch
|
| 8 |
-
from transformers import AutoTokenizer, AutoModel
|
| 9 |
from huggingface_hub import HfApi
|
| 10 |
|
| 11 |
api = HfApi()
|
| 12 |
|
| 13 |
-
#
|
| 14 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
"instruction": [
|
| 16 |
-
("google/flan-t5-small", "
|
| 17 |
-
("google/flan-t5-base", "
|
| 18 |
-
("google-t5/t5-small", "
|
| 19 |
],
|
| 20 |
"qa": [
|
| 21 |
-
("distilbert/distilbert-base-cased-distilled-squad", "
|
| 22 |
-
("distilbert/distilbert-base-uncased-distilled-squad", "
|
| 23 |
-
("deepset/bert-base-cased-squad2", "SQuAD2;
|
| 24 |
],
|
| 25 |
"embeddings": [
|
| 26 |
-
("sentence-transformers/all-MiniLM-L6-v2", "
|
| 27 |
-
("intfloat/e5-small-v2", "
|
| 28 |
-
("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2", "Multilingual
|
| 29 |
],
|
| 30 |
}
|
| 31 |
|
|
|
|
|
|
|
|
|
|
| 32 |
# -----------------------
|
| 33 |
-
#
|
| 34 |
# -----------------------
|
| 35 |
_HUB_CACHE: Dict[Tuple[str, str], Tuple[float, List[str]]] = {}
|
| 36 |
-
CACHE_TTL_SEC = 6 * 60 * 60 #
|
| 37 |
-
|
| 38 |
-
def _language_tag_predicate(tags: List[str],
|
| 39 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
return True
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
tags_lower = {t.lower() for t in (tags or [])}
|
| 45 |
return any(c in tags_lower for c in candidates)
|
| 46 |
|
| 47 |
-
def hub_bonus_models(pipeline_tag: str,
|
| 48 |
-
key = (pipeline_tag,
|
| 49 |
now = time.time()
|
| 50 |
|
| 51 |
if key in _HUB_CACHE:
|
|
@@ -54,13 +133,12 @@ def hub_bonus_models(pipeline_tag: str, language: str, limit: int = 12) -> List[
|
|
| 54 |
return cached
|
| 55 |
|
| 56 |
try:
|
| 57 |
-
# list_models: filtrujemy po pipeline tagu i sortujemy po pobraniach (popularność).
|
| 58 |
models = api.list_models(filter=pipeline_tag, sort="downloads", direction=-1, limit=limit)
|
| 59 |
out = []
|
| 60 |
for m in models:
|
| 61 |
mid = getattr(m, "modelId", None)
|
| 62 |
tags = getattr(m, "tags", []) or []
|
| 63 |
-
if mid and _language_tag_predicate(tags,
|
| 64 |
out.append(mid)
|
| 65 |
_HUB_CACHE[key] = (now, out)
|
| 66 |
return out
|
|
@@ -68,187 +146,154 @@ def hub_bonus_models(pipeline_tag: str, language: str, limit: int = 12) -> List[
|
|
| 68 |
return []
|
| 69 |
|
| 70 |
# -----------------------
|
| 71 |
-
#
|
| 72 |
# -----------------------
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
def
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
batch = texts[i:i+batch_size]
|
| 91 |
-
enc = tok(batch, padding=True, truncation=True, return_tensors="pt")
|
| 92 |
-
out = mdl(**enc)
|
| 93 |
-
|
| 94 |
-
# Mean pooling po tokenach z maską attention
|
| 95 |
-
token_emb = out.last_hidden_state # [B, T, H]
|
| 96 |
-
mask = enc["attention_mask"].unsqueeze(-1).expand(token_emb.size()).float()
|
| 97 |
-
summed = torch.sum(token_emb * mask, dim=1)
|
| 98 |
-
counts = torch.clamp(mask.sum(dim=1), min=1e-9)
|
| 99 |
-
mean_pooled = summed / counts
|
| 100 |
-
|
| 101 |
-
# Normalizacja L2 pomaga dla cosine similarity
|
| 102 |
-
normed = torch.nn.functional.normalize(mean_pooled, p=2, dim=1)
|
| 103 |
-
all_vecs.append(normed.cpu().numpy())
|
| 104 |
-
|
| 105 |
-
return np.vstack(all_vecs)
|
| 106 |
-
|
| 107 |
-
def deduplicate_notes(model_id: str, raw_notes: str, threshold: float) -> str:
|
| 108 |
-
notes = [n.strip() for n in raw_notes.splitlines() if n.strip()]
|
| 109 |
-
if len(notes) < 2:
|
| 110 |
-
return "Wklej co najmniej 2 wpisy (po jednej linijce)."
|
| 111 |
-
|
| 112 |
-
vecs = embed_texts(model_id, notes)
|
| 113 |
-
sim = cosine_similarity(vecs)
|
| 114 |
-
|
| 115 |
-
# Grupowanie prostym union-find (spójne składowe przy sim >= threshold)
|
| 116 |
-
parent = list(range(len(notes)))
|
| 117 |
-
|
| 118 |
-
def find(x):
|
| 119 |
-
while parent[x] != x:
|
| 120 |
-
parent[x] = parent[parent[x]]
|
| 121 |
-
x = parent[x]
|
| 122 |
-
return x
|
| 123 |
-
|
| 124 |
-
def union(a, b):
|
| 125 |
-
ra, rb = find(a), find(b)
|
| 126 |
-
if ra != rb:
|
| 127 |
-
parent[rb] = ra
|
| 128 |
-
|
| 129 |
-
for i in range(len(notes)):
|
| 130 |
-
for j in range(i + 1, len(notes)):
|
| 131 |
-
if sim[i, j] >= threshold:
|
| 132 |
-
union(i, j)
|
| 133 |
-
|
| 134 |
-
groups: Dict[int, List[int]] = {}
|
| 135 |
-
for idx in range(len(notes)):
|
| 136 |
-
r = find(idx)
|
| 137 |
-
groups.setdefault(r, []).append(idx)
|
| 138 |
-
|
| 139 |
-
# Interesują nas grupy z duplikatami (rozmiar > 1)
|
| 140 |
-
dup_groups = [g for g in groups.values() if len(g) > 1]
|
| 141 |
-
dup_groups.sort(key=len, reverse=True)
|
| 142 |
-
|
| 143 |
-
if not dup_groups:
|
| 144 |
-
return f"Brak duplikatów przy progu {threshold:.2f}."
|
| 145 |
-
|
| 146 |
-
lines = []
|
| 147 |
-
lines.append(f"Znalezione grupy podobnych wpisów (próg {threshold:.2f}):")
|
| 148 |
-
lines.append("")
|
| 149 |
-
for gi, g in enumerate(dup_groups, start=1):
|
| 150 |
-
lines.append(f"Grupa {gi} (rozmiar {len(g)}):")
|
| 151 |
-
for idx in g:
|
| 152 |
-
lines.append(f"- {notes[idx]}")
|
| 153 |
-
lines.append("Sugestia: zostaw 1 wpis, pozostałe oznacz jako duplikaty.")
|
| 154 |
-
lines.append("")
|
| 155 |
|
| 156 |
-
|
|
|
|
| 157 |
|
| 158 |
# -----------------------
|
| 159 |
-
#
|
| 160 |
# -----------------------
|
| 161 |
-
def recommend(
|
| 162 |
-
if
|
| 163 |
model_type = "embeddings"
|
| 164 |
why = (
|
| 165 |
-
"
|
| 166 |
-
|
|
|
|
| 167 |
)
|
| 168 |
pipeline_tag = "sentence-similarity"
|
| 169 |
-
|
|
|
|
| 170 |
model_type = "qa"
|
| 171 |
why = (
|
| 172 |
-
"
|
| 173 |
-
|
|
|
|
| 174 |
)
|
| 175 |
pipeline_tag = "question-answering"
|
|
|
|
| 176 |
else:
|
| 177 |
model_type = "instruction"
|
| 178 |
why = (
|
| 179 |
-
"
|
| 180 |
-
|
|
|
|
| 181 |
)
|
| 182 |
pipeline_tag = "text-generation"
|
|
|
|
| 183 |
|
| 184 |
recs = RECOMMENDATIONS[model_type].copy()
|
| 185 |
|
| 186 |
-
#
|
| 187 |
-
bonus = hub_bonus_models(pipeline_tag,
|
| 188 |
existing = {mid for mid, _ in recs}
|
| 189 |
bonus = [m for m in bonus if m not in existing]
|
| 190 |
-
|
| 191 |
-
# Dodajemy do 2 bonusów, żeby nie zalać użytkownika
|
| 192 |
for m in bonus[:2]:
|
| 193 |
-
recs.append((m,
|
| 194 |
|
| 195 |
-
lines = []
|
| 196 |
-
lines.append(
|
| 197 |
lines.append("")
|
| 198 |
-
lines.append("
|
| 199 |
lines.append(f"- {why}")
|
| 200 |
lines.append("")
|
| 201 |
-
lines.append(
|
| 202 |
for mid, note in recs[:5]:
|
| 203 |
lines.append(f"- {mid} — {note}")
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
lines.append("")
|
| 207 |
-
lines.append("Zastosowanie do duplikatów (skrót): embeddingi -> cosine similarity -> próg -> grupy.")
|
| 208 |
-
if language in ["PL", "Mieszany"]:
|
| 209 |
-
lines.append("Wskazówka: preferuj model multilingual przy PL/mix języków.")
|
| 210 |
|
| 211 |
return "\n".join(lines)
|
| 212 |
|
| 213 |
# -----------------------
|
| 214 |
-
# UI
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 215 |
# -----------------------
|
| 216 |
-
with gr.Blocks(title="
|
| 217 |
-
gr.Markdown("#
|
| 218 |
|
| 219 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 220 |
task = gr.Dropdown(
|
| 221 |
-
choices=
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 228 |
)
|
| 229 |
-
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
|
| 234 |
-
btn = gr.Button("Zarekomenduj")
|
| 235 |
-
out = gr.Textbox(lines=18, label="Wynik")
|
| 236 |
-
btn.click(fn=recommend, inputs=[task, has_docs, language, cpu_only, priority], outputs=[out])
|
| 237 |
-
|
| 238 |
-
with gr.Tab("Deduplikacja wpisów (embeddingi)"):
|
| 239 |
-
gr.Markdown(
|
| 240 |
-
"Wklej wpisy (po jednej linijce). Space policzy embeddingi lokalnie na CPU i pogrupuje duplikaty.\n"
|
| 241 |
-
"Uwaga: przy bardzo krótkich, technicznych wpisach warto testować próg w zakresie 0.85–0.95."
|
| 242 |
)
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 247 |
)
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
|
|
|
|
|
|
| 253 |
|
| 254 |
demo.launch()
|
|
|
|
| 1 |
import time
|
| 2 |
+
from typing import Dict, List, Tuple, Any
|
| 3 |
|
| 4 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
from huggingface_hub import HfApi
|
| 6 |
|
| 7 |
api = HfApi()
|
| 8 |
|
| 9 |
+
# -----------------------
|
| 10 |
+
# i18n
|
| 11 |
+
# -----------------------
|
| 12 |
+
I18N: Dict[str, Dict[str, str]] = {
|
| 13 |
+
"EN": {
|
| 14 |
+
"title": "Model Fit Finder (CPU)",
|
| 15 |
+
"intro": (
|
| 16 |
+
"Pick your NLP task and constraints. The Space will recommend an appropriate model type "
|
| 17 |
+
"and list at least 3 concrete Hugging Face models, with short rationale."
|
| 18 |
+
),
|
| 19 |
+
"ui_lang": "UI language",
|
| 20 |
+
"tab_main": "Model advisor",
|
| 21 |
+
"task": "What do you want to do?",
|
| 22 |
+
"has_docs": "Do you have your own documents/text to analyze?",
|
| 23 |
+
"data_lang": "Data language",
|
| 24 |
+
"cpu_only": "CPU only",
|
| 25 |
+
"priority": "Priority",
|
| 26 |
+
"recommend_btn": "Recommend",
|
| 27 |
+
"result": "Result",
|
| 28 |
+
"yes": "Yes",
|
| 29 |
+
"no": "No",
|
| 30 |
+
"en": "EN",
|
| 31 |
+
"pl": "PL",
|
| 32 |
+
"mixed": "Mixed",
|
| 33 |
+
"speed": "Speed",
|
| 34 |
+
"quality": "Quality",
|
| 35 |
+
"task_chat": "Chat / instructions / generation",
|
| 36 |
+
"task_qa": "Answer questions from a document (input text)",
|
| 37 |
+
"task_sim": "Semantic similarity / duplicates / search",
|
| 38 |
+
"rec_type": "Recommended model type: {model_type}",
|
| 39 |
+
"rationale": "Rationale:",
|
| 40 |
+
"models_min3": "Models (min. 3):",
|
| 41 |
+
"emb_note": "Note: embedding models do not generate text; they produce vectors for similarity/search.",
|
| 42 |
+
"qa_note": "Note: extractive QA works best when you provide the relevant context text.",
|
| 43 |
+
"instr_note": "Note: instruction-tuned models follow your prompts; smaller variants are CPU-friendly.",
|
| 44 |
+
"bonus_note": "Popular model from Hub (selected by task tag and downloads).",
|
| 45 |
+
},
|
| 46 |
+
"PL": {
|
| 47 |
+
"title": "Model Fit Finder (CPU)",
|
| 48 |
+
"intro": (
|
| 49 |
+
"Wybierz zadanie NLP i ograniczenia. Space zarekomenduje typ modelu "
|
| 50 |
+
"i pokaże co najmniej 3 konkretne modele z Hugging Face wraz z uzasadnieniem."
|
| 51 |
+
),
|
| 52 |
+
"ui_lang": "Język interfejsu",
|
| 53 |
+
"tab_main": "Doradca modeli",
|
| 54 |
+
"task": "Co chcesz zrobić?",
|
| 55 |
+
"has_docs": "Czy masz własne dokumenty/teksty do analizy?",
|
| 56 |
+
"data_lang": "Język danych",
|
| 57 |
+
"cpu_only": "CPU only",
|
| 58 |
+
"priority": "Priorytet",
|
| 59 |
+
"recommend_btn": "Zarekomenduj",
|
| 60 |
+
"result": "Wynik",
|
| 61 |
+
"yes": "Tak",
|
| 62 |
+
"no": "Nie",
|
| 63 |
+
"en": "EN",
|
| 64 |
+
"pl": "PL",
|
| 65 |
+
"mixed": "Mieszany",
|
| 66 |
+
"speed": "Szybkość",
|
| 67 |
+
"quality": "Jakość",
|
| 68 |
+
"task_chat": "Chat / polecenia / generowanie",
|
| 69 |
+
"task_qa": "Odpowiedzi na pytania z dokumentu (tekst wejściowy)",
|
| 70 |
+
"task_sim": "Semantyczne podobieństwo / duplikaty / wyszukiwanie",
|
| 71 |
+
"rec_type": "Rekomendowany typ modelu: {model_type}",
|
| 72 |
+
"rationale": "Uzasadnienie:",
|
| 73 |
+
"models_min3": "Modele (min. 3):",
|
| 74 |
+
"emb_note": "Uwaga: modele embeddingowe nie generują tekstu; produkują wektory do podobieństwa/wyszukiwania.",
|
| 75 |
+
"qa_note": "Uwaga: QA extractive działa najlepiej, gdy podasz kontekst (tekst źródłowy).",
|
| 76 |
+
"instr_note": "Uwaga: modele instrukcyjne wykonują polecenia; mniejsze warianty są przyjazne dla CPU.",
|
| 77 |
+
"bonus_note": "Popularny model z Hub (dobrany po tagu zadania i pobraniach).",
|
| 78 |
+
},
|
| 79 |
+
}
|
| 80 |
+
|
| 81 |
+
def t(ui_lang: str, key: str) -> str:
|
| 82 |
+
return I18N.get(ui_lang, I18N["EN"]).get(key, I18N["EN"].get(key, key))
|
| 83 |
+
|
| 84 |
+
# -----------------------
|
| 85 |
+
# Stable baseline recommendations (min. 3 per type)
|
| 86 |
+
# -----------------------
|
| 87 |
+
RECOMMENDATIONS: Dict[str, List[Tuple[str, str]]] = {
|
| 88 |
"instruction": [
|
| 89 |
+
("google/flan-t5-small", "Light text2text, good CPU baseline for instruction following."),
|
| 90 |
+
("google/flan-t5-base", "Better quality, slower than small; still workable on CPU."),
|
| 91 |
+
("google-t5/t5-small", "Simple text2text fallback when you want a fast baseline."),
|
| 92 |
],
|
| 93 |
"qa": [
|
| 94 |
+
("distilbert/distilbert-base-cased-distilled-squad", "Fast extractive QA on CPU; classic choice."),
|
| 95 |
+
("distilbert/distilbert-base-uncased-distilled-squad", "Very popular SQuAD QA default."),
|
| 96 |
+
("deepset/bert-base-cased-squad2", "SQuAD2; handles 'no answer' cases better."),
|
| 97 |
],
|
| 98 |
"embeddings": [
|
| 99 |
+
("sentence-transformers/all-MiniLM-L6-v2", "Popular sentence embeddings; fast on CPU."),
|
| 100 |
+
("intfloat/e5-small-v2", "Strong retrieval embeddings; good quality/speed tradeoff."),
|
| 101 |
+
("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2", "Multilingual; better for PL/mixed."),
|
| 102 |
],
|
| 103 |
}
|
| 104 |
|
| 105 |
+
# If you want Polish descriptions here as well, keep EN here and localize notes in output.
|
| 106 |
+
# (Model IDs are universal; notes can be in EN and output can add localized note lines.)
|
| 107 |
+
|
| 108 |
# -----------------------
|
| 109 |
+
# Hub bonus models (cache)
|
| 110 |
# -----------------------
|
| 111 |
_HUB_CACHE: Dict[Tuple[str, str], Tuple[float, List[str]]] = {}
|
| 112 |
+
CACHE_TTL_SEC = 6 * 60 * 60 # 6h
|
| 113 |
+
|
| 114 |
+
def _language_tag_predicate(tags: List[str], data_lang_value: str) -> bool:
|
| 115 |
+
"""
|
| 116 |
+
data_lang_value is one of: EN, PL, MIXED (internal values).
|
| 117 |
+
HF tags aren't perfectly consistent; we do best-effort filtering.
|
| 118 |
+
"""
|
| 119 |
+
if data_lang_value == "MIXED":
|
| 120 |
return True
|
| 121 |
+
target = "en" if data_lang_value == "EN" else "pl"
|
| 122 |
+
candidates = {target, f"language:{target}", f"lang:{target}"}
|
| 123 |
+
tags_lower = {str(x).lower() for x in (tags or [])}
|
|
|
|
| 124 |
return any(c in tags_lower for c in candidates)
|
| 125 |
|
| 126 |
+
def hub_bonus_models(pipeline_tag: str, data_lang_value: str, limit: int = 12) -> List[str]:
|
| 127 |
+
key = (pipeline_tag, data_lang_value)
|
| 128 |
now = time.time()
|
| 129 |
|
| 130 |
if key in _HUB_CACHE:
|
|
|
|
| 133 |
return cached
|
| 134 |
|
| 135 |
try:
|
|
|
|
| 136 |
models = api.list_models(filter=pipeline_tag, sort="downloads", direction=-1, limit=limit)
|
| 137 |
out = []
|
| 138 |
for m in models:
|
| 139 |
mid = getattr(m, "modelId", None)
|
| 140 |
tags = getattr(m, "tags", []) or []
|
| 141 |
+
if mid and _language_tag_predicate(tags, data_lang_value):
|
| 142 |
out.append(mid)
|
| 143 |
_HUB_CACHE[key] = (now, out)
|
| 144 |
return out
|
|
|
|
| 146 |
return []
|
| 147 |
|
| 148 |
# -----------------------
|
| 149 |
+
# Internal "task ids" (do NOT depend on UI language)
|
| 150 |
# -----------------------
|
| 151 |
+
TASK_CHAT = "CHAT"
|
| 152 |
+
TASK_QA = "QA"
|
| 153 |
+
TASK_SIM = "SIM"
|
| 154 |
+
|
| 155 |
+
def task_choices(ui_lang: str) -> List[Tuple[str, str]]:
|
| 156 |
+
"""Return Gradio dropdown choices as (label, value)."""
|
| 157 |
+
return [
|
| 158 |
+
(t(ui_lang, "task_chat"), TASK_CHAT),
|
| 159 |
+
(t(ui_lang, "task_qa"), TASK_QA),
|
| 160 |
+
(t(ui_lang, "task_sim"), TASK_SIM),
|
| 161 |
+
]
|
| 162 |
+
|
| 163 |
+
def yesno_choices(ui_lang: str) -> List[Tuple[str, str]]:
|
| 164 |
+
return [(t(ui_lang, "yes"), "YES"), (t(ui_lang, "no"), "NO")]
|
| 165 |
+
|
| 166 |
+
def data_lang_choices(ui_lang: str) -> List[Tuple[str, str]]:
|
| 167 |
+
return [(t(ui_lang, "en"), "EN"), (t(ui_lang, "pl"), "PL"), (t(ui_lang, "mixed"), "MIXED")]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 168 |
|
| 169 |
+
def priority_choices(ui_lang: str) -> List[Tuple[str, str]]:
|
| 170 |
+
return [(t(ui_lang, "speed"), "SPEED"), (t(ui_lang, "quality"), "QUALITY")]
|
| 171 |
|
| 172 |
# -----------------------
|
| 173 |
+
# Recommendation logic
|
| 174 |
# -----------------------
|
| 175 |
+
def recommend(ui_lang: str, task_id: str, has_docs: str, data_lang_value: str, cpu_only: bool, priority: str) -> str:
|
| 176 |
+
if task_id == TASK_SIM:
|
| 177 |
model_type = "embeddings"
|
| 178 |
why = (
|
| 179 |
+
"You want semantic similarity / duplicate detection / search. Use embeddings + cosine similarity."
|
| 180 |
+
if ui_lang == "EN"
|
| 181 |
+
else "Chcesz podobieństwo semantyczne / duplikaty / wyszukiwanie. Użyj embeddingów + podobieństwa cosinusowego."
|
| 182 |
)
|
| 183 |
pipeline_tag = "sentence-similarity"
|
| 184 |
+
note_key = "emb_note"
|
| 185 |
+
elif task_id == TASK_QA:
|
| 186 |
model_type = "qa"
|
| 187 |
why = (
|
| 188 |
+
"You have a context (document/text) and a question. Extractive QA finds answers in the context."
|
| 189 |
+
if ui_lang == "EN"
|
| 190 |
+
else "Masz kontekst (dokument/tekst) i pytanie. QA extractive znajduje odpowiedź w kontekście."
|
| 191 |
)
|
| 192 |
pipeline_tag = "question-answering"
|
| 193 |
+
note_key = "qa_note"
|
| 194 |
else:
|
| 195 |
model_type = "instruction"
|
| 196 |
why = (
|
| 197 |
+
"You want instruction-following responses (chat/explain/summarize). Instruction-tuned models fit best."
|
| 198 |
+
if ui_lang == "EN"
|
| 199 |
+
else "Chcesz odpowiedzi sterowane poleceniem (chat/wyjaśnianie/streszczanie). Najlepsze są modele instrukcyjne."
|
| 200 |
)
|
| 201 |
pipeline_tag = "text-generation"
|
| 202 |
+
note_key = "instr_note"
|
| 203 |
|
| 204 |
recs = RECOMMENDATIONS[model_type].copy()
|
| 205 |
|
| 206 |
+
# Add 1–2 "bonus" models from Hub, filtered by task tag + best-effort language tags.
|
| 207 |
+
bonus = hub_bonus_models(pipeline_tag, data_lang_value, limit=12)
|
| 208 |
existing = {mid for mid, _ in recs}
|
| 209 |
bonus = [m for m in bonus if m not in existing]
|
|
|
|
|
|
|
| 210 |
for m in bonus[:2]:
|
| 211 |
+
recs.append((m, t(ui_lang, "bonus_note")))
|
| 212 |
|
| 213 |
+
lines: List[str] = []
|
| 214 |
+
lines.append(t(ui_lang, "rec_type").format(model_type=model_type))
|
| 215 |
lines.append("")
|
| 216 |
+
lines.append(t(ui_lang, "rationale"))
|
| 217 |
lines.append(f"- {why}")
|
| 218 |
lines.append("")
|
| 219 |
+
lines.append(t(ui_lang, "models_min3"))
|
| 220 |
for mid, note in recs[:5]:
|
| 221 |
lines.append(f"- {mid} — {note}")
|
| 222 |
+
lines.append("")
|
| 223 |
+
lines.append(t(ui_lang, note_key))
|
|
|
|
|
|
|
|
|
|
|
|
|
| 224 |
|
| 225 |
return "\n".join(lines)
|
| 226 |
|
| 227 |
# -----------------------
|
| 228 |
+
# Dynamic UI language updates
|
| 229 |
+
# -----------------------
|
| 230 |
+
def apply_language(ui_lang: str) -> Tuple[Any, ...]:
|
| 231 |
+
"""
|
| 232 |
+
Returns gr.update objects for all UI text elements that should change when language changes.
|
| 233 |
+
"""
|
| 234 |
+
return (
|
| 235 |
+
gr.update(value=f"# {t(ui_lang, 'title')}\n{t(ui_lang, 'intro')}"), # header_md
|
| 236 |
+
gr.update(label=t(ui_lang, "ui_lang")), # ui_lang radio label (cosmetic)
|
| 237 |
+
gr.update(label=t(ui_lang, "task"), choices=task_choices(ui_lang)), # task dropdown
|
| 238 |
+
gr.update(label=t(ui_lang, "has_docs"), choices=yesno_choices(ui_lang)), # has_docs
|
| 239 |
+
gr.update(label=t(ui_lang, "data_lang"), choices=data_lang_choices(ui_lang)), # data_lang
|
| 240 |
+
gr.update(label=t(ui_lang, "cpu_only")), # cpu_only
|
| 241 |
+
gr.update(label=t(ui_lang, "priority"), choices=priority_choices(ui_lang)), # priority
|
| 242 |
+
gr.update(value=t(ui_lang, "recommend_btn")), # button text
|
| 243 |
+
gr.update(label=t(ui_lang, "result")), # output label
|
| 244 |
+
gr.update(label=t(ui_lang, "tab_main")), # tab label (Gradio may not update tab titles live in all versions)
|
| 245 |
+
)
|
| 246 |
+
|
| 247 |
+
# -----------------------
|
| 248 |
+
# UI
|
| 249 |
# -----------------------
|
| 250 |
+
with gr.Blocks(title=I18N["EN"]["title"]) as demo:
|
| 251 |
+
header_md = gr.Markdown(f"# {t('EN', 'title')}\n{t('EN', 'intro')}")
|
| 252 |
|
| 253 |
+
ui_lang = gr.Radio(
|
| 254 |
+
choices=["EN", "PL"],
|
| 255 |
+
value="EN",
|
| 256 |
+
label=t("EN", "ui_lang"),
|
| 257 |
+
)
|
| 258 |
+
|
| 259 |
+
# Tab title live-update is not guaranteed across Gradio versions; we still keep the label update output.
|
| 260 |
+
with gr.Tab(t("EN", "tab_main")) as tab_main:
|
| 261 |
task = gr.Dropdown(
|
| 262 |
+
choices=task_choices("EN"),
|
| 263 |
+
value=TASK_SIM,
|
| 264 |
+
label=t("EN", "task"),
|
| 265 |
+
)
|
| 266 |
+
has_docs = gr.Radio(
|
| 267 |
+
choices=yesno_choices("EN"),
|
| 268 |
+
value="YES",
|
| 269 |
+
label=t("EN", "has_docs"),
|
| 270 |
+
)
|
| 271 |
+
data_lang = gr.Radio(
|
| 272 |
+
choices=data_lang_choices("EN"),
|
| 273 |
+
value="MIXED",
|
| 274 |
+
label=t("EN", "data_lang"),
|
| 275 |
)
|
| 276 |
+
cpu_only = gr.Checkbox(value=True, label=t("EN", "cpu_only"))
|
| 277 |
+
priority = gr.Radio(
|
| 278 |
+
choices=priority_choices("EN"),
|
| 279 |
+
value="SPEED",
|
| 280 |
+
label=t("EN", "priority"),
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 281 |
)
|
| 282 |
+
|
| 283 |
+
recommend_btn = gr.Button(t("EN", "recommend_btn"))
|
| 284 |
+
out = gr.Textbox(lines=18, label=t("EN", "result"))
|
| 285 |
+
|
| 286 |
+
recommend_btn.click(
|
| 287 |
+
fn=recommend,
|
| 288 |
+
inputs=[ui_lang, task, has_docs, data_lang, cpu_only, priority],
|
| 289 |
+
outputs=[out],
|
| 290 |
)
|
| 291 |
+
|
| 292 |
+
# When UI language changes, update labels + choices.
|
| 293 |
+
ui_lang.change(
|
| 294 |
+
fn=apply_language,
|
| 295 |
+
inputs=[ui_lang],
|
| 296 |
+
outputs=[header_md, ui_lang, task, has_docs, data_lang, cpu_only, priority, recommend_btn, out, tab_main],
|
| 297 |
+
)
|
| 298 |
|
| 299 |
demo.launch()
|