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-4bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use mlx-community/clairic-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-4bit") config = load_config("mlx-community/clairic-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-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-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-4bit" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Hermes Agent
How to use mlx-community/clairic-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-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-4bit
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use mlx-community/clairic-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-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-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: 13,730 Bytes
1a48d6b | 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 | from __future__ import annotations
from collections.abc import Callable
from typing import TYPE_CHECKING, Any
if TYPE_CHECKING:
from .api import PredictRequest
from .backend import EncodedRequest
import hashlib
import math
import re
from copy import deepcopy
from dataclasses import dataclass
from itertools import pairwise
import numpy as np
from PIL import Image
from . import image_table_rows, table_rows
from .schema import MAX_LIST_ITEMS
VIEW_ID = "ragged-table-v1"
TARGET_GATE = "target_is_billed_item"
TARGET_ROW = "target_row"
_BILLED_SCOPE = re.compile(
"\\b(?:line[ _-]?items?|billed\\s+(?:products?|services?|items?|goods))\\b",
re.IGNORECASE,
)
_FINANCIAL_SCOPE = re.compile(
"\\b(?:tax(?:es)?|fees?|surcharges?|payments?)\\b", re.IGNORECASE
)
TARGET_INSTRUCTION = "Image 1 is the complete source document. Image 2 isolates one physical row from image 1. Classify and read only the target row in image 2. Use image 1 to identify the table column labels. A column header is not a billed item. Do not select a neighboring row or count row positions. "
@dataclass(frozen=True)
class RaggedRows:
rows: tuple[tuple[int, int], ...]
def views(self, image: Image.Image) -> tuple[Image.Image, ...]:
return tuple(
(
image.crop(
(0, max(0, top - 3), image.width, min(image.height, bottom + 3))
)
for (top, bottom) in self.rows
)
)
def detect(image: Image.Image) -> RaggedRows | None:
if image_table_rows.detect(image) is not None:
return None
if image.mode != "RGB" or image.width < 100 or image.height < 60:
return None
grey = np.asarray(image.convert("L"))
if np.mean(grey < 245) > 0.22:
return None
ink = grey < 200
candidates = []
for index, (top, bottom) in enumerate(image_table_rows._runs(ink.any(axis=1))):
height = bottom - top
if height < 10:
continue
band = ink[top:bottom]
columns = band.any(axis=0)
separators = []
for left, right in image_table_rows._runs(band.sum(axis=0) >= 0.9 * height):
if (
right - left > max(3, height * 0.18)
or left < 3
or right > image.width - 3
or columns[left - 3 : left].any()
or columns[right : right + 3].any()
):
continue
separators.append((left, right))
if len(separators) >= 2:
candidates.append((index, top, bottom, separators, columns))
if len(candidates) < 2:
return None
if any((right[0] != left[0] + 1 for (left, right) in pairwise(candidates))):
return None
if len({len(row[3]) for row in candidates}) != 1:
return None
if any((row[4][row[3][-1][1] + 3 :].any() for row in candidates)):
return None
heights = np.array([bottom - top for (_, top, bottom, _, _) in candidates])
spacing = np.diff([row[1] for row in candidates])
if np.max(heights) > 1.2 * np.min(heights) or np.max(spacing) > 1.3 * np.min(
spacing
):
return None
return RaggedRows(tuple(((row[1], row[2]) for row in candidates)))
def _eligible(spec: dict[str, Any]) -> bool:
item = spec.get("items")
properties = item.get("properties", {}) if isinstance(item, dict) else {}
(scope, *exclusions) = re.split(
"\\bExclude\\b", spec.get("instructions", ""), maxsplit=1, flags=re.IGNORECASE
)
if exclusions:
if re.search(
"\\b(?:not|never|\\w+n['’]t)(?:\\s+\\w+){0,2}\\s*$", scope, re.IGNORECASE
):
return False
(excluded, *later) = re.split("[.;\\n]", exclusions[0], maxsplit=1)
if (
later
and _FINANCIAL_SCOPE.search(later[0])
or (
_FINANCIAL_SCOPE.search(excluded)
and re.search(
"\\b(?:include|including|keep|retain|except|unless|not|no|none)\\b",
excluded,
re.IGNORECASE,
)
)
):
return False
return (
spec.get("type") == "list"
and isinstance(item, dict)
and (item.get("type") == "object")
and (not table_rows.join_wrapped_descriptions(spec.get("instructions", "")))
and (not spec.get("knowledge"))
and (len(properties) >= 3)
and bool(_BILLED_SCOPE.search(scope))
and (not _FINANCIAL_SCOPE.search(scope))
and all(
(
child.get("type") in ("text", "number") and (not child.get("knowledge"))
for child in properties.values()
)
)
)
def row_fields(spec: dict[str, Any]) -> dict[str, dict[str, Any]]:
wrapped = table_rows.join_wrapped_descriptions(spec.get("instructions", ""))
membership = (
"A logical item can span several printed bands. Include it exactly once, at its first description band. Return false for a later continuation band belonging to that preceding item, even when its numeric cells are printed on the continuation band. Distinct separate source rows remain separate even if they repeat. "
if wrapped
else ""
)
readout = (
"For the target item's first band, use the full document to read its complete wrapped description and cells from its continuation bands. Never borrow from a neighboring logical item. "
if wrapped
else ""
)
return {
TARGET_GATE: {
"type": "boolean",
"nullable": False,
"instructions": TARGET_INSTRUCTION
+ membership
+ "Is the target row a billed product or service row? Return false for the column header, totals, tax, fees, or payment rows. "
+ spec.get("instructions", ""),
},
TARGET_ROW: {
"type": "object",
"nullable": False,
"properties": deepcopy(spec["items"]["properties"]),
"instructions": TARGET_INSTRUCTION
+ readout
+ "Read the complete printed target row as an object. Whether it belongs in the requested list is decided separately. Always return an object with every requested cell. Missing cells and absent columns are null. Do not relabel an Amount column as Unit price. Never compute missing values.",
},
}
def prepare(
request: PredictRequest,
encoded: EncodedRequest,
encode: Callable[..., EncodedRequest],
*,
scale: float = 2,
) -> EncodedRequest:
(record, forecasts, tasks, typed, eos) = encoded
annotations = record["_meta"]["extraction_contract"]["fields"]
parents = [
name
for (name, spec) in request.fields.items()
if _eligible(spec) and annotations.get(name, {}).get("image_row_fields")
]
if not parents or len(request.images) != 1 or request.documents:
return encoded
originals = record.get("_original_images")
image = originals[0] if originals else request.images[0].to_pil()
plan = detect(image)
if plan is None:
return encoded
views = []
for view in plan.views(image):
factor = min(scale, (1600000 / (view.width * view.height)) ** 0.5)
if factor > 1:
view = view.resize(
(int(view.width * factor), int(view.height * factor)),
Image.Resampling.BICUBIC,
)
views.append(view)
salts = dict(record.get("_ar_task_salts", {}))
gates = list(record.get("_source_row_gates", []))
changed = False
for name in parents:
(spec, annotation) = (request.fields[name], annotations[name])
previous = annotation["image_row_fields"]
maximum = spec.get("max_items", MAX_LIST_ITEMS)
if len(views) > maximum + 1 or len(views) > len(previous):
continue
(replacements, pairs) = ([], [])
for index, (view, band) in enumerate(zip(views, plan.rows, strict=True)):
child = previous[index]
digest = hashlib.sha256(child.encode()).hexdigest()[:16]
gate = "ragged_gate_" + digest
while gate in record["fields"]:
gate += "_"
row_request = request.model_copy(update={"fields": row_fields(spec)})
(row_record, _, row_tasks, _, _) = encode(
row_request,
physical_row_view=True,
image_views=lambda view=view: (image, view),
image_view_key=f"{VIEW_ID}:{band}:scale{scale}",
)
assert len(row_tasks) == 2
names = {TARGET_GATE: gate, TARGET_ROW: child}
for source, ids, payload in row_tasks:
target = names[source]
record["fields"][target] = row_record["fields"][source]
record["_ar_constraints"][target] = row_record["_ar_constraints"][
source
]
salts[target] = row_record["_cache_salt"]
annotations[target] = {"parent": name}
replacements.append((target, ids, payload))
gates.append(gate)
pairs.append((gate, child))
tasks = [
task for task in tasks if task[0] not in {name, *previous}
] + replacements
retained = {copy for (_, copy) in pairs}
for child in set(previous) - retained:
record["fields"].pop(child, None)
record["_ar_constraints"].pop(child, None)
annotations.pop(child, None)
salts.pop(child, None)
annotation.pop("image_row_fields")
annotation["ragged_row_fields"] = pairs
annotation["ragged_row_max_items"] = maximum
record["fields"][name].pop("_image_row_inventory", None)
changed = True
if not changed:
return encoded
record.update(
_ar_tasks=tasks,
_ar_task_salts=salts,
_source_row_gates=gates,
_ar_prefix=[],
_ar_group_prefixes=[],
)
record.pop("_ar_media_groups", None)
return (record, forecasts, tasks, typed, eos)
def _score(value: object, *, probability: bool = False) -> bool:
return (
isinstance(value, (int, float))
and (not isinstance(value, bool))
and math.isfinite(value)
and (0 <= value <= 1 if probability else value <= 0)
)
def assemble(
record: dict[str, Any], predictions: dict[str, dict[str, Any]]
) -> dict[str, dict[str, Any]]:
result = {}
for name, annotation in record["_meta"]["extraction_contract"]["fields"].items():
pairs = annotation.get("ragged_row_fields")
if not pairs:
continue
rows = [(predictions[gate], predictions[copy]) for (gate, copy) in pairs]
selected = [(gate, copy) for (gate, copy) in rows if gate.get("value") is True]
complete = all((gate.get("complete", False) for (gate, _) in rows)) and all(
(copy.get("complete", False) for (_, copy) in selected)
)
exclusion_scored = all(
(_score(gate.get("p_excluded"), probability=True) for (gate, _) in rows)
)
relevant = [gate for (gate, _) in rows] + [copy for (_, copy) in selected]
scored = exclusion_scored and all(
(_score(answer.get("log_probability")) for answer in relevant)
)
probability = (
sum((answer["log_probability"] for answer in relevant)) if scored else None
)
if probability is not None and (not math.isfinite(probability)):
(scored, probability) = (False, None)
valid = (
complete
and scored
and all(
(
gate.get("valid", False) and type(gate.get("value")) is bool
for (gate, _) in rows
)
)
and all(
(
copy.get("valid", False) and isinstance(copy.get("value"), dict)
for (_, copy) in selected
)
)
)
within_limit = len(selected) <= annotation["ragged_row_max_items"]
valid &= within_limit
answer = {
"value": [copy["value"] for (_, copy) in selected],
"valid": valid,
"complete": complete,
"confidence": math.exp(probability)
if valid and probability is not None
else 0.0,
"confidence_method": "ar_physical_image_row_eligibility_and_copy_probability",
"confidences": [
math.exp(gate["log_probability"] + copy["log_probability"])
if scored
else 0.0
for (gate, copy) in selected
],
"p_null": math.prod((gate["p_excluded"] for (gate, _) in rows))
if exclusion_scored
else 0.0,
"p_null_method": "product_physical_image_row_exclusion_probabilities"
if exclusion_scored
else "unavailable_physical_image_row_exclusion_probability",
"p_nulls": [
gate["p_excluded"] if exclusion_scored else 0.0
for (gate, _) in selected
],
"spans": [None] * len(selected),
}
if probability is not None:
answer["log_probability"] = probability
if not valid:
answer["failure_reason"] = (
"ragged_row_output_exceeds_max_items"
if not within_limit
else "ragged_row_unscored"
if not scored
else "ragged_row_read_incomplete"
)
result[name] = answer
return result
|