Image-Text-to-Text
Transformers
Safetensors
qwen3_5
clairic
cloudflare
systemone
multimodal
structured-output
information-extraction
custom-code
conversational
Instructions to use Cloudflare/clairic with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Cloudflare/clairic with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="Cloudflare/clairic") 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") model = AutoModelForMultimodalLM.from_pretrained("Cloudflare/clairic", 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 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Cloudflare/clairic" # 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", "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
- SGLang
How to use Cloudflare/clairic 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" \ --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", "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" \ --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", "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 with Docker Model Runner:
docker model run hf.co/Cloudflare/clairic
File size: 8,872 Bytes
f65336d | 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 | from __future__ import annotations
from collections.abc import Callable
from typing import TYPE_CHECKING, Any, cast
if TYPE_CHECKING:
from .api import PredictRequest
from numpy.typing import ArrayLike
from .backend import EncodedRequest
import math
from dataclasses import dataclass
import numpy as np
from PIL import Image
from .schema import MAX_LIST_ITEMS
@dataclass(frozen=True)
class RowTable:
header: tuple[int, int]
rows: tuple[tuple[int, int], ...]
def views(self, image: Image.Image) -> list[Image.Image]:
result = []
for top, bottom in self.rows:
header = image.crop(
(
0,
max(0, self.header[0] - 3),
image.width,
min(image.height, self.header[1] + 3),
)
)
row = image.crop(
(0, max(0, top - 3), image.width, min(image.height, bottom + 3))
)
view = Image.new(
"RGB", (image.width, header.height + row.height + 12), "white"
)
view.paste(header, (0, 0))
view.paste(row, (0, header.height + 12))
result.append(view)
return result
def _runs(values: ArrayLike) -> list[tuple[int, int]]:
padded = np.pad(np.asarray(values, dtype=np.int8), (1, 1))
edges = np.flatnonzero(np.diff(padded))
return cast(
list[tuple[int, int]], [tuple(pair) for pair in edges.reshape(-1, 2).tolist()]
)
def detect(image: Image.Image) -> RowTable | 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
bands = _runs(ink.any(axis=1))
height: float
candidates = []
for index, (top, bottom) in enumerate(bands):
height = bottom - top
if height < 10:
continue
band = ink[top:bottom]
columns = band.any(axis=0)
separators = []
for left, right in _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) / 2)
if len(separators) >= 2:
candidates.append((index, top, bottom, separators))
if len(candidates) < 4:
return None
groups: list[list[tuple[int, int, int, list[float]]]] = []
for candidate in candidates:
if not groups or candidate[0] != groups[-1][-1][0] + 1:
groups.append([])
groups[-1].append(candidate)
if len(groups) != 1:
return None
eligible = []
for group in groups:
if len(group) < 4 or len({len(row[3]) for row in group}) != 1:
continue
(header, body) = (group[0], group[1:])
heights = np.array([bottom - top for (_, top, bottom, _) in group])
height = float(np.median(heights))
positions = np.array([row[3] for row in body])
middle = np.median(positions, axis=0)
spacing = np.diff([row[1] for row in group])
if (
np.max(heights) > 1.2 * np.min(heights)
or np.max(spacing) > 1.3 * np.min(spacing)
or np.max(np.abs(positions - middle)) > height * 1.2
or (abs(header[3][0] - middle[0]) < height * 2)
):
continue
eligible.append(
RowTable((header[1], header[2]), tuple(((row[1], row[2]) for row in body)))
)
return eligible[0] if len(eligible) == 1 else None
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, annotation) in annotations.items()
if annotation.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()
table = detect(image)
if table is None:
return encoded
views = table.views(image)
salts = dict(record.get("_ar_task_salts", {}))
for name in parents:
spec = request.fields[name]
if len(views) > spec.get("max_items", MAX_LIST_ITEMS):
continue
annotation = annotations[name]
previous = annotation["image_row_fields"]
replacements = []
for index, view in enumerate(views):
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,
)
child = previous[index]
view_field = "source_row"
row_spec = {
"type": "object",
"nullable": True,
"properties": spec["items"]["properties"],
"instructions": "Check whether the single data row below the column header matches the requested scope. Return null for an excluded row. Otherwise return the complete printed row. "
+ spec.get("instructions", "")
+ " Missing cells are null. Do not count or select a different row.",
}
row_request = request.model_copy(update={"fields": {view_field: row_spec}})
(row_record, _, row_tasks, _, _) = encode(
row_request,
physical_row_view=True,
image_views=lambda view=view: (view,),
image_view_key=f"table-row-v1:{table.header}:{table.rows[index]}:{scale}",
)
assert len(row_tasks) == 1
record["fields"][child] = row_record["fields"][view_field]
record["_ar_constraints"][child] = row_record["_ar_constraints"][view_field]
salts[child] = row_record["_cache_salt"]
(_, ids, payload) = row_tasks[0]
replacements.append((child, ids, payload))
replaced = {name, *previous}
tasks = [task for task in tasks if task[0] not in replaced] + replacements
retained = {task[0] for task in replacements}
for child in set(previous) - retained:
record["fields"].pop(child)
record["_ar_constraints"].pop(child, None)
annotations.pop(child, None)
annotation["geometric_row_fields"] = [task[0] for task in replacements]
annotation.pop("image_row_fields")
record["fields"][name].pop("_image_row_inventory", None)
if not any(
(annotation.get("geometric_row_fields") for annotation in annotations.values())
):
return encoded
record.update(
_ar_tasks=tasks, _ar_task_salts=salts, _ar_prefix=[], _ar_group_prefixes=[]
)
return (record, forecasts, tasks, typed, eos)
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():
queries = annotation.get("geometric_row_fields")
if not queries:
continue
rows = [predictions[query] for query in queries]
selected = [row for row in rows if row.get("value") is not None]
complete = all((row.get("complete", False) for row in rows))
valid = complete and all((row.get("valid", False) for row in rows))
answer = {
"value": [row["value"] for row in selected],
"complete": complete,
"valid": valid,
"confidence": 0.0,
"confidences": [row.get("confidence", 0.0) for row in selected],
"confidence_method": "ar_physical_image_row_sequence_probability",
"p_null": math.prod((row["p_null"] for row in rows)),
"p_nulls": [row["p_null"] for row in selected],
"p_null_method": "product_physical_image_row_exclusion_probabilities",
"spans": [None] * len(selected),
}
if valid and all(("log_probability" in row for row in rows)):
probability = sum((row["log_probability"] for row in rows))
if math.isfinite(probability):
answer.update(
log_probability=probability,
confidence=math.exp(min(0.0, probability)),
)
result[name] = answer
return result
|