Image-Text-to-Text
MLX
Safetensors
qwen3_5
clairic
cloudflare
systemone
structured-output
information-extraction
multimodal
custom-code
conversational
4-bit precision
Instructions to use mlx-community/clairic-flash-4bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use mlx-community/clairic-flash-4bit with MLX:
# Make sure mlx-vlm is installed # pip install --upgrade mlx-vlm from mlx_vlm import load, generate from mlx_vlm.prompt_utils import apply_chat_template from mlx_vlm.utils import load_config # Load the model model, processor = load("mlx-community/clairic-flash-4bit") config = load_config("mlx-community/clairic-flash-4bit") # Prepare input image = ["http://images.cocodataset.org/val2017/000000039769.jpg"] prompt = "Describe this image." # Apply chat template formatted_prompt = apply_chat_template( processor, config, prompt, num_images=1 ) # Generate output output = generate(model, processor, formatted_prompt, image) print(output) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Pi
How to use mlx-community/clairic-flash-4bit with Pi:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "mlx-community/clairic-flash-4bit"
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "mlx-lm": { "baseUrl": "http://localhost:8080/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "mlx-community/clairic-flash-4bit" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Hermes Agent
How to use mlx-community/clairic-flash-4bit with Hermes Agent:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "mlx-community/clairic-flash-4bit"
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:8080/v1 hermes config set model.default mlx-community/clairic-flash-4bit
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use mlx-community/clairic-flash-4bit with OpenClaw:
Start the MLX server
# Install MLX LM: uv tool install mlx-lm # Start a local OpenAI-compatible server: mlx_lm.server --model "mlx-community/clairic-flash-4bit"
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:8080/v1 \ --custom-model-id "mlx-community/clairic-flash-4bit" \ --custom-provider-id mlx-lm \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
File size: 19,952 Bytes
ae97c5d | 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 | from __future__ import annotations
import csv
import json
import math
import re
from collections.abc import Mapping, Sequence
from typing import Any
from jsonschema import Draft202012Validator, ValidationError
from . import ar, html_rows, printed_numbers
LOCAL_SCOPE_INPUT = (
"The row to classify is provided in the requested field instructions."
)
def _column(value: str) -> str:
return re.sub("[\\W_]+", "", value.casefold())
def _header_keys(cells: Sequence[str], names: Mapping[str, str]) -> list[str]:
keys = [_column(cell) for cell in cells]
return [
"quantity"
if key == "qty" and "qty" not in names and ("quantity" in names)
else key
for key in keys
]
def join_wrapped_descriptions(instructions: str) -> bool:
positive = False
for clause in re.split("[.;\\n]", instructions):
wrapped = re.search(
"\\b(?:wrapped|multiline|multi-line|continuation)\\b", clause, re.IGNORECASE
)
joining = re.search("\\b(?:join|combine|merge)\\b", clause, re.IGNORECASE)
negative = re.search(
"\\b(?:not|no|never|without|separate\\w*|independent\\w*)\\b|\\b\\w+n['’]t\\b",
clause,
re.IGNORECASE,
)
if negative and (wrapped or joining):
return False
positive |= bool(
re.search(
"\\b(?:join|combine)\\b[^.;\\n]*\\b(?:wrapped|multiline|multi-line|continuation)\\b[^.;\\n]*\\b(?:descriptions?|lines?|text)\\b",
clause,
re.IGNORECASE,
)
)
return positive
def _wrapped_rows(
rows: list[dict[str, Any]], properties: dict[str, dict[str, Any]]
) -> list[dict[str, Any]] | None:
texts = [name for (name, spec) in properties.items() if spec["type"] == "text"]
numbers = [name for (name, spec) in properties.items() if spec["type"] == "number"]
if len(texts) != 1 or not numbers:
return rows
text = texts[0]
joined: list[dict[str, Any]] = []
for row in rows:
previous = joined[-1] if joined else None
if (
previous
and previous["values"][text]
and row["values"][text]
and (previous["spans"][text] is not None)
and (row["spans"][text] is not None)
and all((previous["values"][key] is None for key in numbers))
and any((row["values"][key] is not None for key in numbers))
and (
row["line"].find(row["values"][text])
>= previous["line"].find(previous["values"][text]) + 2
)
):
value = previous["values"][text] + " " + row["values"][text]
if not Draft202012Validator(ar.schema(properties[text])).is_valid(value):
return None
joined[-1] = {
"line": previous["line"] + "\n" + row["line"],
"values": {**row["values"], text: value},
"spans": {**row["spans"], text: None},
}
else:
joined.append(row)
return joined
def _source_row(
cells: Sequence[str],
keys: Sequence[str],
line: str,
offset: int,
properties: dict[str, dict[str, Any]],
names: Mapping[str, str],
) -> dict[str, Any] | None:
value: str | int | float | None
(values, spans) = ({}, {})
for key, cell in zip(keys, cells, strict=True):
name = names[key]
spec = properties[name]
if not cell:
value = None
elif spec["type"] == "number":
try:
value = printed_numbers.normalize(cell, spec)
except (ValueError, TypeError, OverflowError, ValidationError):
return None
else:
value = cell
if not Draft202012Validator(ar.schema(spec)).is_valid(value):
return None
values[name] = value
start = line.find(cell) if cell and line.count(cell) == 1 else -1
spans[name] = (
[offset + start, offset + start + len(cell)] if start >= 0 else None
)
for name, spec in properties.items():
if name not in values:
if not Draft202012Validator(ar.schema(spec)).is_valid(None):
return None
(values[name], spans[name]) = (None, None)
return {"line": line.rstrip("\r\n"), "values": values, "spans": spans}
def _html_candidates(
source: str, properties: dict[str, dict[str, Any]], names: Mapping[str, str]
) -> list[dict[str, Any]] | None:
tables = html_rows.tables(source)
if tables is None or len(tables) != 1 or len(tables[0]) < 2:
return None
(header, *body) = tables[0]
keys = [_column(cell) for cell in header["cells"]]
if (
len(keys) < 2
or len(set(keys)) != len(keys)
or any((key not in names for key in keys))
):
return None
rows = []
for row in body:
if len(row["cells"]) != len(keys) or not any(row["cells"]):
return None
parsed = _source_row(
row["cells"], keys, row["line"], row["offset"], properties, names
)
if parsed is None:
return None
for key, span in zip(keys, row["spans"], strict=True):
parsed["spans"][names[key]] = span
rows.append(parsed)
return rows
def candidates(
source: str, properties: dict[str, dict[str, Any]], *, wrapped: bool = False
) -> list[dict[str, Any]] | None:
if any(
(
spec["type"] not in ("text", "number") or spec.get("knowledge")
for spec in properties.values()
)
):
return None
names = {_column(name): name for name in properties}
if len(names) != len(properties):
return None
if re.search("<table\\b", source, re.IGNORECASE):
return _html_candidates(source, properties, names)
lines = source.splitlines(keepends=True)
(offsets, offset) = ([], 0)
for line in lines:
offsets.append(offset)
offset += len(line)
tables = []
for delimiter in ("|", "\t", ",", ";"):
for index, line in enumerate(lines):
if delimiter not in line:
continue
try:
header = next(csv.reader([line], delimiter=delimiter, strict=True))
except csv.Error:
continue
header = [cell.strip() for cell in header]
leading = delimiter == "|" and len(header) > 2 and (not header[0])
trailing = delimiter == "|" and len(header) > 2 and (not header[-1])
if trailing:
header = header[:-1]
if leading:
header = header[1:]
keys = _header_keys(header, names)
if (
len(keys) < 2
or len(set(keys)) != len(keys)
or any((key not in names for key in keys))
):
continue
rows: list[dict[str, Any]] = []
for row_index in range(index + 1, len(lines)):
text = lines[row_index]
if delimiter not in text:
break
try:
cells = next(csv.reader([text], delimiter=delimiter, strict=True))
except csv.Error:
return None
cells = [cell.strip() for cell in cells]
if trailing and cells and (not cells[-1]):
cells = cells[:-1]
if leading and cells and (not cells[0]):
cells = cells[1:]
if len(cells) != len(keys):
return None
if not rows and all(
(re.fullmatch(":?-{3,}:?", cell) for cell in cells)
):
continue
if not any(cells):
return None
row = _source_row(
cells, keys, text, offsets[row_index], properties, names
)
if row is None:
return None
rows.append(row)
tables.append(_wrapped_rows(rows, properties) if wrapped else rows)
return tables[0] if len(tables) == 1 else None
def _scope_instruction(instructions: str) -> str | None:
requested = re.split(
"[.\\n]|\\bExclude\\b", instructions, maxsplit=1, flags=re.IGNORECASE
)[0]
roles = [
bool(re.search(pattern, requested, re.IGNORECASE))
for pattern in (
"\\bfees?\\b",
"\\btax(?:es)?\\b",
"\\b(?:charges?|goods|products?|services?)\\b",
)
]
if sum(roles) != 1:
return None
priority = "Evaluate exclusions first for this exact row; an exclusion overrides a general category match. "
if roles[0]:
return (
priority
+ "Include only a separately printed fee or surcharge. A billed product or service charge, tax, total or payment is not a fee merely because it has a monetary amount. The row label must explicitly establish that it is a fee or surcharge. "
)
if roles[1]:
return (
priority
+ "Include only an explicitly identified tax, such as sales tax, VAT or GST. A product, service charge, fee, total or payment is not a tax merely because it has an amount. "
)
if roles[2]:
return (
priority
+ "Include only billed goods or service charges. A separately identified tax, fee, surcharge, payment, subtotal or total is excluded, even though it is a monetary charge. "
)
return None
def scoped_pairs(source: str, spec: dict[str, Any]) -> list[dict[str, Any]] | None:
instructions = spec.get("instructions", "")
scope_instruction = _scope_instruction(instructions)
properties = spec["items"]["properties"]
if (
len(properties) != 2
or scope_instruction is None
or (not re.match("\\s*Only\\b[^.\\n]*\\brows\\b", instructions, re.IGNORECASE))
or re.search(
"\\b(?:summar\\w*|calculat\\w*|combin\\w*|group\\w*|aggregat\\w*|deduplicat\\w*)\\b|\\buse\\b[^.\\n]*\\bas\\b",
instructions,
re.IGNORECASE,
)
):
return None
texts = [
name for (name, child) in properties.items() if child.get("type") == "text"
]
numbers = [
name for (name, child) in properties.items() if child.get("type") == "number"
]
if (
len(texts) != 1
or len(numbers) != 1
or any((child.get("knowledge") for child in properties.values()))
):
return None
if not re.search(
"\\b(?:exact|verbatim)\\b[^.\\n]*\\b(?:printed|row)\\b[^.\\n]*\\blabel\\b",
properties[texts[0]].get("instructions", ""),
re.IGNORECASE,
):
return None
names = {_column(name): name for name in properties}
if len(names) != len(properties):
return None
keys = [_column(texts[0]), _column(numbers[0])]
tables = []
for delimiter in ("|", "\t", ";"):
(rows, offset) = ([], 0)
for line in source.splitlines(keepends=True):
if delimiter in line:
try:
cells = [
cell.strip()
for cell in next(
csv.reader([line], delimiter=delimiter, strict=True)
)
]
except csv.Error:
return None
if len(cells) != 2 or not all(cells):
return None
row = _source_row(cells, keys, line, offset, properties, names)
if row is None:
return None
row["eligibility_only"] = True
row["scope_instruction"] = scope_instruction
rows.append(row)
offset += len(line)
if len(rows) >= 2:
tables.append(rows)
return tables[0] if len(tables) == 1 else None
def prepare(
source: str,
fields: dict[str, dict[str, Any]],
*,
has_media: bool = False,
allow_scoped_pairs: bool = False,
) -> dict[str, list[dict[str, Any]]]:
if has_media:
return {}
result = {}
for name, spec in fields.items():
item = spec.get("items")
if (
spec.get("type") != "list"
or not isinstance(item, dict)
or item.get("type") != "object"
):
continue
explicit_rows = re.search(
"\\bone\\s+(?:object|entry|record)\\s+(?:per|for\\s+each)\\b[^.\\n]*\\brow\\b",
spec.get("instructions", ""),
re.IGNORECASE,
)
wrapped = join_wrapped_descriptions(spec.get("instructions", ""))
rows = (
candidates(source, item["properties"], wrapped=wrapped)
if explicit_rows
else scoped_pairs(source, spec)
if allow_scoped_pairs
else None
)
if rows and len(rows) <= spec.get("max_items", 32):
missing_columns = [
key
for (key, child) in item["properties"].items()
if child.get("type") == "number"
and all((row["values"][key] is None for row in rows))
]
if missing_columns:
for row in rows:
row["eligibility_only"] = True
result[name] = rows
return result
def membership_instructions(instructions: str) -> str:
if re.search(
"[\\r\\n`\\\"“”]|(?<!\\w)'[^'\\r\\n]+'(?!\\w)|‘[^’]+’|\\b(?:literal|verbatim|quoted|quote|quotation|code|snippet|block|note|label|phrase)\\b",
instructions,
re.IGNORECASE,
):
return instructions
target = "(?:(?:wrapped|multiline|multi-line|continuation)\\s+)?(?:descriptions?(?:\\s+lines?)?|lines?|(?:distinct|separate|duplicate|repeated|physical)\\s+rows?)"
negative_join = re.compile(
"\\s*(?:do\\s+not|don't|never)\\s+(?:join|combine|merge)\\s+"
+ target
+ "(?:\\s+or\\s+(?:join|combine|merge)\\s+"
+ target
+ ")*\\s*[.;]?\\s*",
re.IGNORECASE,
)
clauses = re.split("(?<=[.;])(?=\\s|$)|(?<=\\n)", instructions)
if not any((negative_join.fullmatch(clause) for clause in clauses)):
return instructions
return "".join(
(clause for clause in clauses if not negative_join.fullmatch(clause))
).strip()
def row_spec(
spec: dict[str, Any], row: dict[str, Any], *, eligibility: bool = False
) -> dict[str, Any]:
missing = [name for (name, value) in row["values"].items() if value is None]
missing_instruction = (
"Missing cells or entire columns do not exclude an otherwise matching row. These source cells are missing and must remain null: "
+ ", ".join(missing)
+ ". "
if missing
else ""
)
scope = membership_instructions(spec.get("instructions", ""))
if eligibility:
return {
"type": "boolean",
"nullable": False,
"instructions": "Classify exactly this physical source row for inclusion in the requested list. Return true if and only if this row satisfies the requested list instructions. Return false if the row is excluded, unrelated, or does not belong to the requested list, even though it exists. This is an eligibility decision, not a request to copy its cells. "
+ missing_instruction
+ row.get("scope_instruction", "")
+ "Requested list instructions: "
+ scope
+ "\nExact physical source row: "
+ row["line"],
}
properties = {}
for name, child in spec["items"]["properties"].items():
value = row["values"][name]
properties[name] = {**child, "nullable": value is None, "enum": [value]}
return {
**spec["items"],
"properties": properties,
"nullable": True,
"instructions": scope
+ "\nRead only this exact physical source row: "
+ json.dumps(row["line"], ensure_ascii=False)
+ "\nReturn null if this row does not satisfy the requested scope. Otherwise copy its cells together. "
+ missing_instruction
+ "Do not substitute a different row. Missing cells are null. This query has no list ordinal to count.",
}
def local_scope_context(
rows: Sequence[dict[str, Any]], queries: Sequence[str], *, offset: int = 0
) -> tuple[dict[str, str], dict[str, str]]:
if not rows or not all((row.get("scope_instruction") for row in rows)):
return ({}, {})
pairs = list(zip(queries, rows, strict=True))
inputs = dict.fromkeys(queries, LOCAL_SCOPE_INPUT)
names = {
query: f"source_scope_{offset + index}"
for (index, (query, _)) in enumerate(pairs, 1)
}
return (inputs, names)
def assemble(
predictions: dict[str, dict[str, Any]],
rows: Sequence[dict[str, Any]],
queries: Sequence[str],
*,
eligibility: bool = False,
) -> dict[str, Any]:
checks = [predictions[name] for name in queries]
selected = [
(row, answer)
for (row, answer) in zip(rows, checks, strict=True)
if (
answer.get("value") is True
if eligibility
else answer.get("value") is not None
)
]
valid = all(
(
answer.get("valid", False) and answer.get("complete", False)
for answer in checks
)
)
scored = True
if eligibility:
valid &= all((type(answer.get("value")) is bool for answer in checks))
scored = all(
(
isinstance((score := answer.get("p_excluded")), (int, float))
and (not isinstance(score, bool))
and math.isfinite(score)
and (0 <= score <= 1)
for answer in checks
)
)
valid &= scored
p_null = (
math.exp(sum((math.log(answer["p_excluded"]) for answer in checks)))
if scored and all((answer["p_excluded"] > 0 for answer in checks))
else 0.0
)
p_null_method = (
"product_source_row_exclusion_probabilities"
if scored
else "unavailable_source_row_exclusion_probability"
)
else:
p_null = min((answer.get("p_null", 0.0) for answer in checks))
p_null_method = "minimum_source_row_null_probability"
log_probability = sum(
(answer.get("log_probability", -math.inf) for answer in checks)
)
if not math.isfinite(log_probability):
log_probability = None
result = {
"value": [
dict(row["values"]) if eligibility else answer["value"]
for (row, answer) in selected
],
"valid": valid,
"complete": valid,
"confidence": math.exp(min(0, log_probability))
if valid and log_probability is not None
else 0.0,
"confidence_method": "ar_source_row_eligibility_probability"
if eligibility
else "ar_source_row_sequence_probability",
"p_null": p_null,
"p_null_method": p_null_method,
"confidences": [answer.get("confidence", 0.0) for (row, answer) in selected],
"p_nulls": [
answer.get("p_excluded" if eligibility else "p_null", 0.0)
for (row, answer) in selected
],
"spans": [None] * len(selected),
"property_spans": [row["spans"] for (row, answer) in selected],
}
if log_probability is not None:
result["log_probability"] = log_probability
if eligibility and (not scored):
result.pop("p_nulls")
if not valid:
result["failure_reason"] = (
"source_row_eligibility_unscored"
if eligibility and (not scored)
else "source_row_read_incomplete"
)
return result
|