Image-Text-to-Text
Transformers
Safetensors
qwen3_5
clairic
cloudflare
systemone
multimodal
structured-output
information-extraction
custom-code
conversational
Instructions to use Cloudflare/clairic-flash with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Cloudflare/clairic-flash with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="Cloudflare/clairic-flash") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForMultimodalLM processor = AutoProcessor.from_pretrained("Cloudflare/clairic-flash") model = AutoModelForMultimodalLM.from_pretrained("Cloudflare/clairic-flash", device_map="auto") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use Cloudflare/clairic-flash with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Cloudflare/clairic-flash" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Cloudflare/clairic-flash", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/Cloudflare/clairic-flash
- SGLang
How to use Cloudflare/clairic-flash with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "Cloudflare/clairic-flash" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Cloudflare/clairic-flash", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "Cloudflare/clairic-flash" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Cloudflare/clairic-flash", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }' - Docker Model Runner
How to use Cloudflare/clairic-flash with Docker Model Runner:
docker model run hf.co/Cloudflare/clairic-flash
File size: 12,610 Bytes
dc7d8aa | 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 | from __future__ import annotations
import json
import math
import re
import unicodedata
from collections.abc import Callable
from typing import Any, cast
from . import MAX_OUTPUT_TOKENS
from .schema import MAX_LIST_ITEMS
INVENTORY_OUTPUT_TOKENS = 2048
_ROW_REQUEST = re.compile(
"\\bone\\s+(?:object|entry|record)\\s+(?:per|for\\s+each)\\b[^.\\n]*\\brow\\b",
re.IGNORECASE,
)
LABEL = {
"type": "text",
"nullable": False,
"instructions": "Brief exact printed identifier of this requested row, without secondary description or numeric cells. Follow the requested field scope: the label may name a product, service, tax, fee, payment or other row.",
}
def prepare(
record: dict[str, Any], *, has_media: bool, share_description: bool = False
) -> dict[str, Any]:
if not has_media:
return record
annotations = record["_meta"]["extraction_contract"]["fields"]
for name, spec in record["fields"].items():
annotation = annotations.get(name, {})
if (
spec.get("type") != "list"
or not isinstance(spec.get("items"), dict)
or spec["items"].get("type") != "object"
or (not annotation.get("query_fields"))
or annotation.get("source_fields")
or (not _ROW_REQUEST.search(spec.get("instructions", "")))
):
continue
spec.update(nullable=True, _image_row_inventory=True)
annotation["image_row_fields"] = annotation.pop("query_fields")
for index, child in enumerate(annotation["image_row_fields"], 1):
row = record["fields"][child]
row.update(nullable=False, _image_row_proposal=True)
if share_description:
row["_share_image_row_description"] = True
row["instructions"] = (
f"Item {index} of {name} in the supplied document. "
+ spec.get("instructions", "")
+ " Return null for each missing cell, including absent entire columns. If this physical row does not exist return an object with every cell null."
)
return record
def is_inventory(spec: dict[str, Any]) -> bool:
return bool(spec.get("_image_row_inventory"))
def inventory_schema(spec: dict[str, Any]) -> dict[str, Any]:
return {
"type": ["array", "null"],
"minItems": 1,
"maxItems": spec.get("max_items", MAX_LIST_ITEMS) + 1,
"items": {
"type": "object",
"properties": {"label": {"type": "string"}},
"required": ["label"],
"additionalProperties": False,
},
}
def instruction(
record: dict[str, Any],
name: str,
schema: Callable[[dict[str, Any]], dict[str, Any]],
) -> str | None:
spec = record["fields"][name]
if is_inventory(spec):
return (
"Extract the requested value from the supplied input, following the field instructions and scope. Return all matching rows together as one JSON array. Return null only if there are no matching rows. Each object describes one physical row. Missing cells are null inside that object; they do not make the row absent. Keep wrapped text in its row and preserve distinct duplicate rows. Stop after the last matching row and close the array. End immediately after the array or null."
+ "\nField: "
+ name
+ "\nSchema: "
+ json.dumps(
inventory_schema(spec), ensure_ascii=False, separators=(",", ":")
)
+ "\nInstructions: "
+ spec.get("instructions", "")
+ "\nProperty instructions: "
+ json.dumps({"label": LABEL}, ensure_ascii=False)
)
if spec.get("_image_row_proposal"):
if spec.get("_share_image_row_description"):
shared = shared_description_instruction(record, name, schema)
if shared is not None:
return shared
return (
"Extract one complete requested row from the supplied document. Return one JSON object matching the schema. Keep all properties on the same physical row. Use null for a missing cell in an existing row. Preserve distinct rows with identical values. Follow the requested row scope and exclusions. End immediately after the object."
+ "\nField: "
+ name
+ "\nSchema: "
+ json.dumps(schema(spec), ensure_ascii=False, separators=(",", ":"))
+ "\nInstructions: "
+ spec.get("instructions", "")
+ "\nProperty instructions: "
+ json.dumps(spec["properties"], ensure_ascii=False)
)
return None
def shared_description_instruction(
record: dict[str, Any],
name: str,
schema: Callable[[dict[str, Any]], dict[str, Any]],
) -> str | None:
spec = record["fields"][name]
if not spec.get("_image_row_proposal"):
return None
annotation = (
record.get("_meta", {})
.get("extraction_contract", {})
.get("fields", {})
.get(name, {})
)
(parent, index) = (annotation.get("parent"), annotation.get("index"))
if (
not isinstance(parent, str)
or not parent
or type(index) is not int
or (index < 1)
):
return None
properties = spec.get("properties", {})
description = properties.get("description", {})
if description.get("type") != "text" or not description.get("instructions"):
return None
ordinal = f"Item {index} of {parent} in the supplied document. "
scope = spec.get("instructions", "")
if not scope.startswith(ordinal):
return None
return (
"Extract one complete requested row from the supplied document. Return one JSON object matching the schema. Keep all properties on the same physical row. Use null for a missing cell in an existing row. Preserve distinct rows with identical values. Follow the requested row scope and exclusions. End immediately after the object."
+ "\nSchema: "
+ json.dumps(schema(spec), ensure_ascii=False, separators=(",", ":"))
+ "\nInstructions: "
+ scope[len(ordinal) :]
+ "\nProperty instructions: "
+ json.dumps(properties, ensure_ascii=False)
+ "\nRequested row: "
+ ordinal.rstrip()
+ "\nField: "
+ name
+ "\nProperty instructions: "
+ json.dumps({"description": description}, ensure_ascii=False)
)
def output_limit(record: dict[str, Any], name: str, length: int, max_input: int) -> int:
desired = (
INVENTORY_OUTPUT_TOKENS
if is_inventory(record["fields"][name])
else MAX_OUTPUT_TOKENS
)
return min(desired, max_input + MAX_OUTPUT_TOKENS - length)
def _words(value: str) -> str:
return " ".join(re.findall("\\w+", unicodedata.normalize("NFKC", value).casefold()))
def _assignment(costs: list[list[float]]) -> list[int]:
if not costs:
return []
(count, width) = (len(costs), len(costs[0]))
(u, v, p, way) = (
[0.0] * (count + 1),
[0.0] * (width + 1),
[0] * (width + 1),
[0] * (width + 1),
)
for index in range(1, count + 1):
(p[0], column) = (index, 0)
(minimum, used) = ([math.inf] * (width + 1), [False] * (width + 1))
while True:
used[column] = True
(row, delta, next_column) = (p[column], math.inf, 0)
for candidate in range(1, width + 1):
if not used[candidate]:
cost = costs[row - 1][candidate - 1] - u[row] - v[candidate]
if cost < minimum[candidate]:
(minimum[candidate], way[candidate]) = (cost, column)
if minimum[candidate] < delta:
(delta, next_column) = (minimum[candidate], candidate)
for candidate in range(width + 1):
if used[candidate]:
u[p[candidate]] += delta
v[candidate] -= delta
else:
minimum[candidate] -= delta
column = next_column
if p[column] == 0:
break
while column:
previous = way[column]
(p[column], column) = (p[previous], previous)
result: list[int | None] = [None] * count
for column in range(1, width + 1):
if p[column]:
result[p[column] - 1] = column - 1
return cast(list[int], result)
def _select(
labels: list[str], proposals: list[dict[str, Any]], text_keys: list[str]
) -> tuple[list[dict[str, Any] | None], list[int]]:
if not text_keys:
return (
[
proposals[index] if index < len(proposals) else None
for index in range(len(labels))
],
[],
)
costs = []
for index, label in enumerate(labels):
label = _words(label)
row_costs = []
for position, proposal in enumerate(proposals):
value = proposal.get("value") if proposal.get("valid") else None
texts = (
[value.get(key) for key in text_keys] if isinstance(value, dict) else []
)
matches = label and any(
(
isinstance(text, str) and f" {label} " in f" {_words(text)} "
for text in texts
)
)
row_costs.append(abs(position - index) * 10 if matches else 1000000000.0)
costs.append(row_costs + [100000000.0] * len(labels))
chosen = _assignment(costs)
conflicts = [
index
for (index, column) in enumerate(chosen)
if column >= len(proposals) or costs[index][column] >= 100000000.0
]
return (
[
None if index in conflicts else proposals[column]
for (index, column) in enumerate(chosen)
],
conflicts,
)
def assemble(
record: dict[str, Any], predictions: dict[str, dict[str, Any]]
) -> dict[str, dict[str, Any]]:
answers = {}
annotations = record["_meta"]["extraction_contract"]["fields"]
for name, annotation in annotations.items():
queries = annotation.get("image_row_fields")
if not queries:
continue
(spec, inventory) = (record["fields"][name], predictions[name])
labels = [row["label"] for row in inventory.get("value") or []]
maximum = spec.get("max_items", MAX_LIST_ITEMS)
inventory_complete = (
inventory.get("valid", False)
and inventory.get("complete", False)
and (len(labels) <= maximum)
)
properties = spec["items"]["properties"]
text_keys = [
key for (key, child) in properties.items() if child["type"] == "text"
]
(selected, conflicts) = _select(
labels[:maximum], [predictions[child] for child in queries], text_keys
)
complete = inventory_complete and all(
(row and row.get("complete", False) for row in selected)
)
valid = (
complete
and (not conflicts)
and all(
(
row
and row.get("valid")
and isinstance(row.get("value"), dict)
and any((value is not None for value in row["value"].values()))
for row in selected
)
)
)
answer = {
"value": [row["value"] if row else None for row in selected],
"valid": valid,
"complete": complete,
"confidence": 0.0,
"confidences": [
row.get("confidence", 0.0) if row else 0.0 for row in selected
],
"confidence_method": "ar_sequence_probability_with_source_row_labels",
"p_null": inventory["p_null"],
"p_null_method": "source_row_labels_null_probability",
"spans": [None] * len(selected),
}
scores = [inventory, *selected]
if valid and all((row and "log_probability" in row for row in scores)):
probability = sum(
(row["log_probability"] for row in cast(list[dict[str, Any]], scores))
)
if math.isfinite(probability):
answer.update(
log_probability=probability,
confidence=math.exp(min(0.0, probability)),
)
if conflicts:
answer["failure_reason"] = "source_row_alignment"
answers[name] = answer
return answers
|