Spaces:
Runtime error
Runtime error
File size: 23,685 Bytes
bf96836 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 | """
Tokenizer Inspector module.
Provides tokenization utilities and a Gradio UI tab for inspecting how
different tokenizers handle input text.
"""
import html
import gradio as gr
from langdetect import detect, LangDetectException
_AutoTokenizer = None
def _get_auto_tokenizer():
global _AutoTokenizer
if _AutoTokenizer is None:
from transformers import AutoTokenizer
_AutoTokenizer = AutoTokenizer
return _AutoTokenizer
class _LazyAutoTokenizer:
"""Proxy that defers transformers import until first attribute access."""
def __getattr__(self, name):
return getattr(_get_auto_tokenizer(), name)
AutoTokenizer = _LazyAutoTokenizer()
from openrouter import call_openrouter
# ---------------------------------------------------------------------------
# Tokenizer registry
# ---------------------------------------------------------------------------
SUPPORTED_TOKENIZERS: dict[str, str] = {
"o200k_base": "tiktoken:o200k_base",
"cl100k_base": "tiktoken:cl100k_base",
"llama-3": "NousResearch/Meta-Llama-3-8B",
"mistral": "mistralai/Mistral-7B-v0.1",
"qwen-2.5": "Qwen/Qwen2.5-7B",
"gemma-2": "microsoft/phi-2",
"command-r": "bigscience/bloom-560m",
"gpt2": "gpt2",
}
class TiktokenAdapter:
"""Wraps a tiktoken encoding to match the HuggingFace tokenizer interface.
This allows tokenize_text(), fragmentation_ratio(), and other functions
to work identically with tiktoken-based and HF-based tokenizers.
"""
def __init__(self, encoding_name: str):
import tiktoken
self._enc = tiktoken.get_encoding(encoding_name)
self._encoding_name = encoding_name
def encode(self, text: str, add_special_tokens: bool = True) -> list[int]:
return self._enc.encode(text)
def decode(self, token_ids: list[int]) -> str:
return self._enc.decode(token_ids)
def convert_ids_to_tokens(self, token_ids: list[int]) -> list[str]:
return [self._enc.decode([tid]) for tid in token_ids]
def __repr__(self) -> str:
return f"TiktokenAdapter({self._encoding_name!r})"
# Module-level cache: name -> tokenizer object
_tokenizer_cache: dict[str, object] = {}
def get_tokenizer(name: str):
"""Return (and cache) a tokenizer for the given registry name.
Args:
name: Key in SUPPORTED_TOKENIZERS (e.g. 'o200k_base', 'llama-3').
Returns:
A loaded tokenizer (AutoTokenizer or TiktokenAdapter).
Raises:
ValueError: If name is not in SUPPORTED_TOKENIZERS.
"""
if name not in SUPPORTED_TOKENIZERS:
raise ValueError(f"unknown tokenizer: '{name}'. Choose from {list(SUPPORTED_TOKENIZERS)}")
if name not in _tokenizer_cache:
repo_id = SUPPORTED_TOKENIZERS[name]
if repo_id.startswith("tiktoken:"):
encoding_name = repo_id.split(":", 1)[1]
_tokenizer_cache[name] = TiktokenAdapter(encoding_name)
else:
try:
_tokenizer_cache[name] = AutoTokenizer.from_pretrained(repo_id)
except Exception as exc:
raise RuntimeError(
f"Failed to load tokenizer '{name}' from '{repo_id}'. "
f"Check your network connection or set TRANSFORMERS_OFFLINE=1 "
f"if you have a local cache. Original error: {exc}"
) from exc
return _tokenizer_cache[name]
# ---------------------------------------------------------------------------
# Core tokenization helpers
# ---------------------------------------------------------------------------
def tokenize_text(text: str, tokenizer) -> list[dict]:
"""Tokenize text and return a list of {token, id} dicts.
Args:
text: Input string to tokenize.
tokenizer: A loaded AutoTokenizer (or compatible mock).
Returns:
List of dicts with keys 'token' (str) and 'id' (int).
"""
token_ids = tokenizer.encode(text)
tokens = tokenizer.convert_ids_to_tokens(token_ids)
return [{"token": str(tok), "id": int(tid)} for tok, tid in zip(tokens, token_ids)]
def fragmentation_ratio(text: str, tokenizer) -> dict[str, float]:
"""Compute the fragmentation ratio (tokens per word) for text.
Args:
text: Input string.
tokenizer: A loaded AutoTokenizer.
Returns:
Dict with:
- 'ratio': float tokens-per-word (0.0 when text is empty)
- 'token_count': int total token count
"""
token_ids = tokenizer.encode(text)
token_count = len(token_ids)
words = text.split()
word_count = len(words)
ratio = token_count / word_count if word_count > 0 else 0.0
return {"ratio": float(ratio), "token_count": token_count}
def flag_oov_words(text: str, tokenizer, threshold: int = 3) -> set[str]:
"""Return the set of words that fragment into >= threshold tokens.
A word is considered out-of-vocabulary (OOV) relative to a tokenizer when
the tokenizer splits it into many sub-word pieces.
Args:
text: Input string.
tokenizer: A loaded AutoTokenizer.
threshold: Minimum token count (inclusive) to flag a word. Default 3.
Returns:
Set of words that meet or exceed the threshold.
"""
oov: set[str] = set()
for word in text.split():
ids = tokenizer.encode(word, add_special_tokens=False)
if len(ids) >= threshold:
oov.add(word)
return oov
# ---------------------------------------------------------------------------
# Language helpers
# ---------------------------------------------------------------------------
def detect_language(text: str) -> str:
"""Detect the language of text.
Wraps langdetect.detect. Returns 'en' on LangDetectException.
Args:
text: Input string.
Returns:
BCP-47 language code string, e.g. 'en', 'fr', 'de'.
"""
try:
return detect(text)
except LangDetectException:
return "en"
def efficiency_score(input_tokens: int, english_tokens: int) -> float:
"""Compute tokenization efficiency relative to an English translation.
Score = english_tokens / input_tokens. Values > 1.0 indicate the source
language is more compact than English for this tokenizer; < 1.0 means
more tokens are needed.
Args:
input_tokens: Token count for the original (possibly non-English) text.
english_tokens: Token count for the English translation.
Returns:
Float ratio. Returns 1.0 when english_tokens is 0 or input_tokens is 0.
"""
if english_tokens == 0 or input_tokens == 0:
return 1.0
return float(english_tokens) / float(input_tokens)
# ---------------------------------------------------------------------------
# Token tax metrics (GH-3)
# ---------------------------------------------------------------------------
def relative_tokenization_cost(source_tokens: int, english_tokens: int) -> float:
"""Relative Tokenization Cost: source_tokens / english_tokens.
Values > 1.0 indicate the source language pays a "token tax" vs English.
Values < 1.0 indicate the source is more compact than English.
Args:
source_tokens: Token count for the (possibly non-English) text.
english_tokens: Token count for the English equivalent.
Returns:
Float ratio. Returns 1.0 when english_tokens is 0 (zero guard).
"""
if english_tokens == 0:
return 1.0
return float(source_tokens) / float(english_tokens)
def byte_premium(text: str, english_text: str) -> float:
"""Ratio of UTF-8 byte length of text vs english_text.
Values > 1.0 indicate the source text uses more bytes than English
for equivalent content, reflecting script-level overhead.
Args:
text: Source text.
english_text: English equivalent text.
Returns:
Float ratio. Returns 1.0 when english_text is empty (zero guard).
"""
source_bytes = len(text.encode("utf-8"))
english_bytes = len(english_text.encode("utf-8"))
if english_bytes == 0:
return 1.0
return float(source_bytes) / float(english_bytes)
def context_window_usage(token_count: int, window_size: int = 128_000) -> float:
"""Fraction of a context window consumed by a token count.
Args:
token_count: Number of tokens.
window_size: Total context window size. Default 128k.
Returns:
Float between 0.0 and 1.0+. Returns 1.0 when window_size is 0.
"""
if window_size == 0:
return 1.0
return float(token_count) / float(window_size)
def quality_risk_level(rtc: float) -> str:
"""Map a Relative Tokenization Cost to a quality risk band.
Based on multilingual tokenization research (2025-2026):
- low (<1.5): tokenizer handles this language well
- moderate (1.5-2.5): noticeable token inflation
- high (2.5-4.0): significant cost and potential quality impact
- severe (>= 4.0): extreme fragmentation, likely quality degradation
Args:
rtc: Relative Tokenization Cost value.
Returns:
One of "low", "moderate", "high", "severe".
"""
if rtc < 1.5:
return "low"
if rtc < 2.5:
return "moderate"
if rtc < 4.0:
return "high"
return "severe"
def translate_to_english(text: str, api_key: str) -> str:
"""Translate text to English using OpenRouter.
Args:
text: Source text to translate.
api_key: OpenRouter API key.
Returns:
Translated English string.
"""
prompt = (
f"Translate the following text to English. "
f"Return only the translation, no explanations.\n\nText: {text}"
)
response = call_openrouter(api_key, "meta-llama/llama-3.1-8b-instruct", prompt)
return response["choices"][0]["message"]["content"]
# ---------------------------------------------------------------------------
# HTML rendering
# ---------------------------------------------------------------------------
_NORMAL_BG_COLOURS = ("#e8f4f8", "#d4ecd4")
def render_tokens_html(
tokens: list[dict],
oov_words: set[str],
tokenizer=None,
decoded_view: bool = False,
hide_special_tokens: bool = True,
) -> str:
"""Render a list of tokens as coloured HTML spans.
Normal tokens alternate between two background colours. OOV tokens are
highlighted with #ffcccc. All token text is HTML-escaped.
When decoded_view=True, token IDs are decoded back into readable text
snippets. Special tokens (e.g. BOS) are hidden by default in this mode.
"""
special_ids = set(getattr(tokenizer, "all_special_ids", [])) if tokenizer else set()
parts: list[str] = []
display_chunks: list[str] = []
token_ids = [int(entry["id"]) for entry in tokens]
byte_decoder = getattr(tokenizer, "byte_decoder", None) if tokenizer else None
has_convert_tokens_to_string = bool(tokenizer and hasattr(tokenizer, "convert_tokens_to_string"))
if decoded_view and tokenizer is not None:
used_convert_path = False
# Prefer tokenizer-native token-string reconstruction when available.
# This handles byte-level tokenizers more reliably than per-token decode.
if has_convert_tokens_to_string:
tmp_chunks: list[str] = []
visible_tokens: list[str] = []
prev_decoded = ""
convert_path_ok = True
for i, token_id in enumerate(token_ids):
if hide_special_tokens and token_id in special_ids:
tmp_chunks.append("")
continue
visible_tokens.append(str(tokens[i]["token"]))
try:
curr_decoded = tokenizer.convert_tokens_to_string(visible_tokens)
except Exception:
convert_path_ok = False
break
if not isinstance(curr_decoded, str):
convert_path_ok = False
break
safe_prev = prev_decoded.replace("\ufffd", "")
safe_curr = curr_decoded.replace("\ufffd", "")
chunk = safe_curr[len(safe_prev):] if safe_curr.startswith(safe_prev) else ""
tmp_chunks.append(chunk)
prev_decoded = curr_decoded
if convert_path_ok:
display_chunks = tmp_chunks
used_convert_path = True
# For byte-level tokenizers (GPT-2/Llama-family), decode via raw byte
# accumulation to avoid replacement-character noise in multibyte scripts.
if not used_convert_path and isinstance(byte_decoder, dict) and byte_decoder:
buffer = bytearray()
prev_decoded = ""
for i, token_id in enumerate(token_ids):
if hide_special_tokens and token_id in special_ids:
display_chunks.append("")
continue
raw_token = str(tokens[i]["token"])
for ch in raw_token:
if ch in byte_decoder:
buffer.append(int(byte_decoder[ch]))
else:
buffer.extend(ch.encode("utf-8", errors="ignore"))
curr_decoded = bytes(buffer).decode("utf-8", errors="ignore")
chunk = curr_decoded[len(prev_decoded):] if curr_decoded.startswith(prev_decoded) else ""
display_chunks.append(chunk)
prev_decoded = curr_decoded
elif not used_convert_path:
# Generic fallback: cumulative tokenizer decode + prefix diff.
prev_decoded = ""
for idx, token_id in enumerate(token_ids):
if hide_special_tokens and token_id in special_ids:
display_chunks.append("")
continue
try:
curr_decoded = tokenizer.decode(
token_ids[: idx + 1],
skip_special_tokens=hide_special_tokens,
clean_up_tokenization_spaces=False,
)
except Exception:
curr_decoded = prev_decoded
if curr_decoded.startswith(prev_decoded):
chunk = curr_decoded[len(prev_decoded):]
else:
# Conservative fallback if tokenizer decode is non-prefix-stable.
try:
chunk = tokenizer.decode(
[token_id],
skip_special_tokens=hide_special_tokens,
clean_up_tokenization_spaces=False,
)
except Exception:
chunk = ""
# Strip replacement chars in readable mode.
if "\ufffd" in chunk:
chunk = chunk.replace("\ufffd", "")
display_chunks.append(chunk)
prev_decoded = curr_decoded
else:
display_chunks = [str(entry["token"]) for entry in tokens]
for i, entry in enumerate(tokens):
raw_token = str(entry["token"])
token_id = int(entry["id"])
token_text_for_display = display_chunks[i]
if decoded_view and tokenizer is not None:
if not token_text_for_display and hide_special_tokens:
continue
token_text = html.escape(token_text_for_display)
if raw_token in oov_words:
bg = "#ffcccc"
else:
bg = _NORMAL_BG_COLOURS[i % 2]
title = f"id:{token_id}" if decoded_view else f"id:{token_id} | raw:{html.escape(raw_token)}"
parts.append(
f'<span style="background:{bg};padding:2px 4px;border-radius:3px;'
f'margin:1px;display:inline-block;white-space:pre;color:#000;" title="{title}">'
f"{token_text}</span>"
)
return "".join(parts)
# ---------------------------------------------------------------------------
# Gradio UI
# ---------------------------------------------------------------------------
def _handle_single(
model_name: str,
text: str,
threshold: int,
decoded_view: bool,
english_text: str = "",
):
"""Handler logic for the Single tab — extracted for testability."""
try:
tok = get_tokenizer(model_name)
tokens = tokenize_text(text, tok)
token_count = len(tokens)
oov = flag_oov_words(text, tok, threshold=int(threshold))
token_html = render_tokens_html(
tokens, oov, tokenizer=tok,
decoded_view=decoded_view, hide_special_tokens=True,
)
frag = fragmentation_ratio(text, tok)
lang = detect_language(text)
ctx_usage = context_window_usage(token_count, 128_000)
stats = (
f"**Tokens:** {frag['token_count']} \n"
f"**Fragmentation ratio:** {frag['ratio']:.2f} \n"
f"**OOV words:** {len(oov)} \n"
f"**Detected language:** {lang} \n"
f"**Context usage (128k):** {ctx_usage:.4%}"
)
if lang == "en":
stats += (
f" \n**RTC vs English:** 1.0x \n"
f"**Quality risk:** low"
)
elif english_text and english_text.strip():
eng_tokens = tokenize_text(english_text.strip(), tok)
rtc = relative_tokenization_cost(token_count, len(eng_tokens))
risk = quality_risk_level(rtc)
stats += (
f" \n**RTC vs English:** {rtc:.2f}x \n"
f"**Quality risk:** {risk}"
)
else:
stats += (
" \n**RTC:** *(provide English equivalent for comparison)*"
)
return token_html, stats
except Exception as exc:
return "", f"Error: {exc}"
def _handle_compare(
text: str,
name_a: str,
name_b: str,
decoded_view: bool,
english_text: str = "",
):
"""Handler logic for the Compare tab — extracted for testability."""
try:
tok_a = get_tokenizer(name_a)
tok_b = get_tokenizer(name_b)
tokens_a = tokenize_text(text, tok_a)
tokens_b = tokenize_text(text, tok_b)
count_a = len(tokens_a)
count_b = len(tokens_b)
html_a = render_tokens_html(
tokens_a, set(), tokenizer=tok_a,
decoded_view=decoded_view, hide_special_tokens=True,
)
html_b = render_tokens_html(
tokens_b, set(), tokenizer=tok_b,
decoded_view=decoded_view, hide_special_tokens=True,
)
frag_a = fragmentation_ratio(text, tok_a)
frag_b = fragmentation_ratio(text, tok_b)
ratio_md = (
f"**{name_a}:** {frag_a['token_count']} tokens "
f"(ratio {frag_a['ratio']:.2f}) \n"
f"**{name_b}:** {frag_b['token_count']} tokens "
f"(ratio {frag_b['ratio']:.2f})"
)
if english_text and english_text.strip():
eng_tokens_a = tokenize_text(english_text.strip(), tok_a)
eng_tokens_b = tokenize_text(english_text.strip(), tok_b)
rtc_a = relative_tokenization_cost(count_a, len(eng_tokens_a))
rtc_b = relative_tokenization_cost(count_b, len(eng_tokens_b))
ratio_md += (
f" \n**{name_a} RTC:** {rtc_a:.2f}x \n"
f"**{name_b} RTC:** {rtc_b:.2f}x"
)
if rtc_a != rtc_b:
better = name_a if rtc_a < rtc_b else name_b
ratio_md += (
f" \n*{better} is more efficient for this language.*"
)
return html_a, html_b, ratio_md
except Exception as exc:
return "", "", f"Error: {exc}"
def build_tokenizer_ui() -> gr.Blocks:
"""Construct and return the Tokenizer Inspector Gradio Blocks UI.
Returns:
gr.Blocks instance with two inner tabs:
- Single: inspect tokenization of one model.
- Compare: side-by-side comparison of two models.
"""
tokenizer_names = list(SUPPORTED_TOKENIZERS.keys())
with gr.Blocks(title="Tokenizer Inspector") as demo:
gr.Markdown("## Tokenizer Inspector\nExplore how different tokenizers split text.")
with gr.Tabs():
# --- Single tab ---
with gr.TabItem("Single"):
with gr.Row():
single_model = gr.Dropdown(
choices=tokenizer_names,
value=tokenizer_names[0],
label="Tokenizer",
)
single_text = gr.Textbox(
label="Input Text",
placeholder="Type text to tokenize...",
lines=3,
)
oov_threshold = gr.Slider(
minimum=1, maximum=10, value=3, step=1,
label="OOV threshold (tokens per word)",
)
single_english_text = gr.Textbox(
label="English Equivalent (optional)",
placeholder="Paste English translation for RTC comparison...",
lines=2,
)
single_decoded_view = gr.Checkbox(
label="Readable token view (decode tokens, hide special tokens)",
value=False,
)
single_btn = gr.Button("Tokenize", variant="primary")
single_html = gr.HTML(label="Token Visualisation")
single_stats = gr.Markdown(label="Statistics")
single_btn.click(
fn=_handle_single,
inputs=[single_model, single_text, oov_threshold, single_decoded_view, single_english_text],
outputs=[single_html, single_stats],
)
# --- Compare tab ---
with gr.TabItem("Compare"):
compare_text = gr.Textbox(
label="Input Text",
placeholder="Type text to compare tokenizers...",
lines=3,
)
with gr.Row():
cmp_model_a = gr.Dropdown(
choices=tokenizer_names,
value=tokenizer_names[0],
label="Tokenizer A",
)
cmp_model_b = gr.Dropdown(
choices=tokenizer_names,
value=tokenizer_names[1] if len(tokenizer_names) > 1 else tokenizer_names[0],
label="Tokenizer B",
)
compare_english_text = gr.Textbox(
label="English Equivalent (optional)",
placeholder="Paste English translation for RTC comparison...",
lines=2,
)
compare_btn = gr.Button("Compare", variant="primary")
with gr.Row():
cmp_html_a = gr.HTML(label="Tokenizer A")
cmp_html_b = gr.HTML(label="Tokenizer B")
compare_decoded_view = gr.Checkbox(
label="Readable token view (decode tokens, hide special tokens)",
value=False,
)
cmp_ratio_md = gr.Markdown(label="Comparison")
compare_btn.click(
fn=_handle_compare,
inputs=[compare_text, cmp_model_a, cmp_model_b, compare_decoded_view, compare_english_text],
outputs=[cmp_html_a, cmp_html_b, cmp_ratio_md],
)
return demo
|