AutoRAG 0.3.25__tar.gz → 0.3.26__tar.gz

autorag-0.3.26/AGENTS.md

@@ -0,0 +1,14 @@
+# Legacy AutoRAG contribution guide
+
+## Type checking
+
+`ty` is required for legacy product code only. Do not run `ty` against
+`legacy/tests`, and do not add test files to the `ty` command or CI scope.
+
+Tests are validated with pytest and ruff. Test doubles often intentionally
+implement only the runtime surface needed by a test and are therefore outside
+the legacy product-code typing contract.
+
+The current clean product scope is documented in
+`legacy/docs/source/ty.md`. Extend that scope with product modules as they are
+made clean; keep tests excluded.

{autorag-0.3.25 → autorag-0.3.26}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: AutoRAG
-Version: 0.3.25
+Version: 0.3.26
 Summary: Automatically Evaluate RAG pipelines with your own data. Find optimal structure for new RAG product.
 Project-URL: Homepage, https://github.com/Marker-Inc-Korea/AutoRAG
 Author-email: Marker-Inc <vkehfdl1@gmail.com>
@@ -229,10 +229,10 @@ Requires-Dist: evaluate>=0.4.4
 Requires-Dist: fastapi>=0.115.13
 Requires-Dist: fastparquet>=2024.11.0
 Requires-Dist: gradio>=6.15.0
-Requires-Dist: grpcio-health-checking<1.84.0,>=1.66.2
-Requires-Dist: grpcio-status<1.84.0,>=1.66.2
-Requires-Dist: grpcio-tools<1.84.0,>=1.66.2
-Requires-Dist: grpcio<1.84.0,>=1.66.2
+Requires-Dist: grpcio-health-checking<1.85.0,>=1.66.2
+Requires-Dist: grpcio-status<1.85.0,>=1.66.2
+Requires-Dist: grpcio-tools<1.85.0,>=1.66.2
+Requires-Dist: grpcio<1.85.0,>=1.66.2
 Requires-Dist: ipykernel>=6.29.5
 Requires-Dist: ipywidgets-bokeh>=1.6.0
 Requires-Dist: ipywidgets>=8.1.7
@@ -301,7 +301,7 @@ Requires-Dist: sudachipy>=0.6.8; extra == 'all'
 Requires-Dist: torch>=2.7.1; extra == 'all'
 Requires-Dist: transformers>=4.51.3; extra == 'all'
 Requires-Dist: unstructured[pdf]>=0.17.2; extra == 'all'
-Requires-Dist: vllm<0.28.0,>=0.11.0; extra == 'all'
+Requires-Dist: vllm<0.30.0,>=0.11.0; extra == 'all'
 Requires-Dist: xgrammar<0.3.0,>=0.1.32; extra == 'all'
 Provides-Extra: gpu
 Requires-Dist: bert-score>=0.3.13; extra == 'gpu'
@@ -316,7 +316,6 @@ Requires-Dist: sentence-transformers>=4.1.0; extra == 'gpu'
 Requires-Dist: sentencepiece>=0.2.0; extra == 'gpu'
 Requires-Dist: torch>=2.7.1; extra == 'gpu'
 Requires-Dist: transformers>=4.51.3; extra == 'gpu'
-Requires-Dist: vllm<0.28.0,>=0.11.0; extra == 'gpu'
 Requires-Dist: xgrammar<0.3.0,>=0.1.32; extra == 'gpu'
 Provides-Extra: ja
 Requires-Dist: sudachidict-core; extra == 'ja'
@@ -332,6 +331,8 @@ Requires-Dist: pdfplumber>=0.11.7; extra == 'parse'
 Requires-Dist: pymupdf>=1.26.1; extra == 'parse'
 Requires-Dist: pypdf2>=3.0.1; extra == 'parse'
 Requires-Dist: unstructured[pdf]>=0.17.2; extra == 'parse'
+Provides-Extra: vllm
+Requires-Dist: vllm<0.30.0,>=0.11.0; extra == 'vllm'
 Description-Content-Type: text/markdown
 
 # AutoRAG (Legacy)

{autorag-0.3.25 → autorag-0.3.26}/autorag/data/chunk/langchain_chunk.py

@@ -31,6 +31,7 @@ def langchain_chunk(
 	:param metadata_list: The list of dict of metadata from the parsed result
 	:return: tuple of lists containing the chunked doc_id, contents, path, start_idx, end_idx and metadata
 	"""
+	metadata_list = metadata_list or [{} for _ in texts]
 	results = [
 		langchain_chunk_pure(text, chunker, file_name_language, meta)
 		for text, meta in zip(texts, metadata_list)
@@ -50,7 +51,9 @@ def langchain_chunk_pure(
 	_metadata: Optional[Dict[str, str]] = None,
 ):
 	# chunk
-	chunk_results = chunker.create_documents([text], metadatas=[_metadata])
+	chunk_results = chunker.create_documents(
+		[text], metadatas=[_metadata or {}]
+	)
 
 	# make doc_id
 	doc_id = list(str(uuid.uuid4()) for _ in range(len(chunk_results)))

{autorag-0.3.25 → autorag-0.3.26}/autorag/data/chunk/llama_index_chunk.py

@@ -37,6 +37,7 @@ def llama_index_chunk(
 	:param batch: The batch size for chunk texts. Default is 8
 	:return: tuple of lists containing the chunked doc_id, contents, path, start_idx, end_idx and metadata
 	"""
+	metadata_list = metadata_list or [{} for _ in texts]
 	tasks = [
 		llama_index_chunk_pure(text, chunker, file_name_language, meta)
 		for text, meta in zip(texts, metadata_list)
@@ -58,7 +59,7 @@ async def llama_index_chunk_pure(
 	_metadata: Optional[Dict[str, str]] = None,
 ):
 	# set document
-	document = [Document(text=text, metadata=_metadata)]
+	document = [Document(text=text, metadata=_metadata or {})]
 
 	# chunk document
 	chunk_results = await chunker.aget_nodes_from_documents(documents=document)
@@ -72,7 +73,7 @@ async def llama_index_chunk_pure(
 	# make contents and start_end_idx
 	if file_name_language:
 		chunked_file_names = list(map(lambda x: os.path.basename(x), path_lst))
-		chunked_texts = list(map(lambda x: x.text, chunk_results))
+		chunked_texts = list(map(lambda x: x.get_content(), chunk_results))
 		start_end_idx = list(
 			map(
 				lambda x: get_start_end_idx(text, x),
@@ -81,7 +82,7 @@ async def llama_index_chunk_pure(
 		)
 		contents = add_file_name(file_name_language, chunked_file_names, chunked_texts)
 	else:
-		contents = list(map(lambda x: x.text, chunk_results))
+		contents = list(map(lambda x: x.get_content(), chunk_results))
 		start_end_idx = list(map(lambda x: get_start_end_idx(text, x), contents))
 
 	metadata = list(

{autorag-0.3.25 → autorag-0.3.26}/autorag/data/parse/base.py

@@ -48,6 +48,8 @@ def parser_node(func):
 		)
 
 		if func.__name__ == "langchain_parse":
+			if parse_method is None:
+				raise ValueError("parse_method is required for langchain_parse")
 			parse_method = parse_method.lower()
 			if parse_method == "directory":
 				path_split_list = data_path_glob.split("/")

{autorag-0.3.25 → autorag-0.3.26}/autorag/data/parse/clova.py

@@ -2,7 +2,7 @@ import base64
 import itertools
 import json
 import os
-from typing import List, Optional, Tuple
+from typing import List, Optional, Tuple, TypedDict
 
 import aiohttp
 import fitz  # PyMuPDF
@@ -11,6 +11,12 @@ from autorag.data.parse.base import parser_node
 from autorag.utils.util import process_batch, get_event_loop
 
 
+class TableCell(TypedDict):
+	text: str
+	rowSpan: int
+	colSpan: int
+
+
 @parser_node
 def clova_ocr(
 	data_path_list: List[str],
@@ -152,7 +158,9 @@ def json_to_html_table(json_data):
 	max_row = max(cell["rowIndex"] + cell["rowSpan"] for cell in json_data)
 	max_col = max(cell["columnIndex"] + cell["columnSpan"] for cell in json_data)
 	# Create a 2D array to keep track of merged cells
-	table = [["" for _ in range(max_col)] for _ in range(max_row)]
+	table: list[list[TableCell | None | str]] = [
+		["" for _ in range(max_col)] for _ in range(max_row)
+	]
 	# Fill the table with cell data
 	for cell in json_data:
 		row = cell["rowIndex"]
@@ -182,6 +190,8 @@ def json_to_html_table(json_data):
 			if cell == "":
 				html += "    <td></td>\n"
 			else:
+				if not isinstance(cell, dict):
+					raise TypeError(f"Unexpected table cell value: {cell!r}")
 				row_span_attr = (
 					f' rowspan="{cell["rowSpan"]}"' if cell["rowSpan"] > 1 else ""
 				)

{autorag-0.3.25 → autorag-0.3.26}/autorag/nodes/passagereranker/tart/modeling_enc_t5.py

@@ -1,14 +1,34 @@
 # Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
 
 import copy
+from math import ceil
 
 from transformers.modeling_outputs import SequenceClassifierOutput
 from transformers.models.t5.modeling_t5 import T5Config, T5PreTrainedModel, T5Stack
-from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
 
 from autorag.utils.util import empty_cuda_cache
 
 
+def assert_device_map(device_map, num_blocks):
+	blocks = list(range(num_blocks))
+	device_map_blocks = [block for blocks in device_map.values() for block in blocks]
+	duplicate_blocks = [block for block in device_map_blocks if device_map_blocks.count(block) > 1]
+	missing_blocks = [block for block in blocks if block not in device_map_blocks]
+	extra_blocks = [block for block in device_map_blocks if block not in blocks]
+	if duplicate_blocks or missing_blocks or extra_blocks:
+		raise ValueError("device_map must contain each encoder block exactly once")
+
+
+def get_device_map(n_layers, devices):
+	layers = list(range(n_layers))
+	layers_per_device = int(ceil(n_layers / len(devices)))
+	layers_list = [
+		layers[index : index + layers_per_device]
+		for index in range(0, n_layers, layers_per_device)
+	]
+	return dict(zip(devices, layers_list))
+
+
 class EncT5ForSequenceClassification(T5PreTrainedModel):
 	_keys_to_ignore_on_load_missing = [
 		r"encoder\.embed_tokens\.weight",
@@ -28,7 +48,8 @@ class EncT5ForSequenceClassification(T5PreTrainedModel):
 		encoder_config = copy.deepcopy(config)
 		encoder_config.use_cache = False
 		encoder_config.is_encoder_decoder = False
-		self.encoder = T5Stack(encoder_config, self.shared)
+		self.encoder = T5Stack(encoder_config)
+		self.encoder.embed_tokens = self.shared
 
 		self.dropout = nn.Dropout(dropout)
 		self.classifier = nn.Linear(config.hidden_size, config.num_labels)

{autorag-0.3.25 → autorag-0.3.26}/autorag/nodes/passagereranker/tart/tart.py

@@ -8,6 +8,7 @@ from autorag.nodes.passagereranker.tart.modeling_enc_t5 import (
 	EncT5ForSequenceClassification,
 )
 from autorag.nodes.passagereranker.tart.tokenization_enc_t5 import EncT5Tokenizer
+from transformers import T5Config
 from autorag.utils.util import (
 	make_batch,
 	sort_by_scores,
@@ -28,7 +29,15 @@ class Tart(BasePassageReranker):
 				"torch is not installed. Please install torch first to use TART reranker."
 			)
 		model_name = "facebook/tart-full-flan-t5-xl"
-		self.model = EncT5ForSequenceClassification.from_pretrained(model_name)
+		config_dict, _ = T5Config.get_config_dict(model_name)
+		config_dict["id2label"] = {
+			int(key): str(value) for key, value in config_dict["id2label"].items()
+		}
+		config = T5Config.from_dict(config_dict)
+		self.model = EncT5ForSequenceClassification.from_pretrained(
+			model_name,
+			config=config,
+		)
 		self.tokenizer = EncT5Tokenizer.from_pretrained(model_name)
 		self.device = "cuda" if torch.cuda.is_available() else "cpu"
 		self.model = self.model.to(self.device)

{autorag-0.3.25 → autorag-0.3.26}/autorag/nodes/util.py

@@ -4,6 +4,7 @@ from autorag.support import get_support_modules
 
 
 def make_generator_callable_param(generator_dict: Optional[Dict]):
+	generator_dict = generator_dict or {}
 	if "generator_module_type" not in generator_dict.keys():
 		generator_dict = {
 			"generator_module_type": "llama_index_llm",

{autorag-0.3.25 → autorag-0.3.26}/docs/source/index.rst

@@ -83,6 +83,7 @@ Also, feel free to ask your question at our `github issue <https://github.com/Ma
    azure_openai.md
    migration.md
    test_your_rag.md
+   ty.md
 
 .. toctree::
     :maxdepth: 1

autorag-0.3.26/docs/source/ty.md

@@ -0,0 +1,30 @@
+# Legacy AutoRAG type checking
+
+Legacy AutoRAG uses `ty` for static type checking. The checks are intentionally
+introduced in small product-code scopes so optional integrations and third-party
+stub limitations do not hide errors in the core modules.
+
+`legacy/tests` is intentionally excluded from `ty`. Tests and test doubles are
+validated with pytest and ruff instead; they are not part of the product-code
+typing contract.
+
+## Reproducible commands
+
+From `legacy/`:
+
+```bash
+uv sync --extra ko --extra parse --extra ja
+uv run --locked ty check autorag/data/chunk autorag/data/parse/base.py \
+  autorag/data/parse/clova.py autorag/nodes/util.py
+uv run --locked ruff check autorag tests
+uv run --locked pytest -q tests/autorag/data/chunk \
+  tests/autorag/data/parse/test_langchain_parse.py \
+  tests/autorag/data/parse/test_clova.py
+```
+
+The focused product-code `ty` scope above is clean and runs in CI. Do not
+replace it with `ty check autorag tests`: tests are intentionally excluded.
+The complete product-code command, `uv run --locked ty check autorag`, remains
+broader than this initial scope because it includes optional integrations and
+third-party library APIs that are being migrated in subsequent issue-sized
+changes.

{autorag-0.3.25 → autorag-0.3.26}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "AutoRAG"
-version = "0.3.25"
+version = "0.3.26"
 authors = [
     { name = "Marker-Inc", email = "vkehfdl1@gmail.com" }
 ]
@@ -46,10 +46,10 @@ dependencies = [
     "emoji>=2.14.1",
     "fastapi>=0.115.13",
     "banks>=2.4.2",
-    "grpcio>=1.66.2,<1.84.0",
-    "grpcio-health-checking>=1.66.2,<1.84.0",
-    "grpcio-status>=1.66.2,<1.84.0",
-    "grpcio-tools>=1.66.2,<1.84.0",
+    "grpcio>=1.66.2,<1.85.0",
+    "grpcio-health-checking>=1.66.2,<1.85.0",
+    "grpcio-status>=1.66.2,<1.85.0",
+    "grpcio-tools>=1.66.2,<1.85.0",
     "datasets>=3.5.1",
     # Vector DB
     "pymilvus>=2.6.0b0,<4",  # for using milvus vectordb (3.x drops milvus-lite)
@@ -120,7 +120,6 @@ dev = [
     "pytest-mock",
     "aioresponses>=0.7.9",
     "asyncstdlib",
-    "vllm-mock>=0.0.3; sys_platform == 'linux'",
     "ty",
     "deptry",
 ]
@@ -131,9 +130,9 @@ parse = ["PyMuPDF>=1.26.1", "pdfminer.six>=20250506", "pdfplumber>=0.11.7", "jq>
 ja = ["sudachipy>=0.6.8", "sudachidict_core"]
 gpu = ["torch>=2.7.1", "sentencepiece>=0.2.0", "bert_score>=0.3.13", "peft>=0.15.2", "llmlingua>=0.2.2", "FlagEmbedding>=1.2.11",
     "sentence-transformers>=4.1.0", "transformers>=4.51.3", "llama-index-llms-ollama>=0.6.0", "llama-index-embeddings-huggingface>=0.5.4",
-    "llama-index-llms-huggingface>=0.5.0", "onnxruntime>=1.22.0", "xgrammar>=0.1.32,<0.3.0",
-    "vllm>=0.11.0,<0.28.0"]
-all = ["AutoRAG[gpu]", "AutoRAG[ko]", "AutoRAG[parse]", "AutoRAG[ja]"]
+    "llama-index-llms-huggingface>=0.5.0", "onnxruntime>=1.22.0", "xgrammar>=0.1.32,<0.3.0"]
+vllm = ["vllm>=0.11.0,<0.30.0"]
+all = ["AutoRAG[gpu]", "AutoRAG[ko]", "AutoRAG[parse]", "AutoRAG[ja]", "AutoRAG[vllm]"]
 
 [project.scripts]
 autorag = "autorag.cli:cli"

{autorag-0.3.25 → autorag-0.3.26}/tests/autorag/data/parse/test_langchain_parse.py

@@ -1,3 +1,5 @@
+import pytest
+
 from autorag.data.parse import langchain_parse
 
 from tests.autorag.data.parse.test_parse_base import (
@@ -30,6 +32,16 @@ def test_langchain_parse_single_pdf():
     assert pages == [-1]
 
 
+def test_langchain_parse_requires_parse_method():
+    with pytest.raises(ValueError, match="parse_method is required"):
+        langchain_parse(
+            single_pdf_path_list[0].replace(
+                single_pdf_path_list[0].split("/")[-1], "*"
+            ),
+            file_type="pdf",
+        )
+
+
 def test_langchain_parse_single_pdf_node():
     result_df = langchain_parse(
         korean_text_glob, file_type="pdf", parse_method="pdfminer"

{autorag-0.3.25 → autorag-0.3.26}/tests/autorag/nodes/generator/test_vllm.py

@@ -1,14 +1,11 @@
-import pytest
-import pandas as pd
+import sys
+import types
+from dataclasses import dataclass
+from types import SimpleNamespace
 
-pytest.importorskip("vllm", reason="vllm not installed")
-try:
-	import vllm_mock
-	from vllm_mock import LLM
-except Exception as exc:  # noqa: BLE001 - skip when mock is incompatible with installed vllm
-	pytest.skip(f"vllm_mock not compatible with installed vllm: {exc}", allow_module_level=True)
+import pandas as pd
 
-from autorag.nodes.generator import Vllm  # noqa: E402
+from autorag.nodes.generator import Vllm
 from tests.autorag.nodes.generator.test_generator_base import (  # noqa: E402
 	prompts,
 	chat_prompts,
@@ -20,9 +17,68 @@ from tests.autorag.nodes.generator.test_generator_base import (  # noqa: E402
 	check_generated_log_probs_chat,
 )
 
-def test_vllm(mocker):
-	mock_class = mocker.patch("vllm.LLM")
-	mock_class.return_value = LLM(model="mock-model")
+
+@dataclass
+class FakeLogProb:
+	logprob: float
+
+
+class FakeSamplingParams:
+	@classmethod
+	def from_optional(cls, **kwargs):
+		return kwargs
+
+	def __init__(self, **kwargs):
+		self.kwargs = kwargs
+
+
+class FakeLLM:
+	def __init__(self, model, **kwargs):
+		self.model = model
+		self.kwargs = kwargs
+
+	def _outputs(self, prompts):
+		return [
+			SimpleNamespace(
+				outputs=[
+					SimpleNamespace(
+						text=f"generated: {prompt}",
+						token_ids=[1, 2, 3],
+						logprobs=[
+							{1: FakeLogProb(-0.1)},
+							{2: FakeLogProb(-0.2)},
+							{3: FakeLogProb(-0.3)},
+						],
+					)
+				]
+			)
+			for prompt in prompts
+		]
+
+	def generate(self, prompts, sampling_params, **kwargs):
+		return self._outputs(prompts)
+
+	def chat(self, prompts, sampling_params, **kwargs):
+		return self._outputs(
+			[message["content"] for prompt in prompts for message in prompt[-1:]]
+		)
+
+
+def install_fake_vllm(monkeypatch):
+	fake_vllm = types.ModuleType("vllm")
+	fake_outputs = types.ModuleType("vllm.outputs")
+	fake_logprobs = types.ModuleType("vllm.logprobs")
+	setattr(fake_vllm, "LLM", FakeLLM)
+	setattr(fake_vllm, "SamplingParams", FakeSamplingParams)
+	setattr(fake_outputs, "RequestOutput", SimpleNamespace)
+	setattr(fake_logprobs, "SampleLogprobs", dict)
+	monkeypatch.setitem(sys.modules, "vllm", fake_vllm)
+	monkeypatch.setitem(sys.modules, "vllm.outputs", fake_outputs)
+	monkeypatch.setitem(sys.modules, "vllm.logprobs", fake_logprobs)
+
+
+def test_vllm(monkeypatch):
+	install_fake_vllm(monkeypatch)
 
 	previous_result = pd.DataFrame(
 		{"prompts": prompts, "qid": ["id-1", "id-2", "id-3"]}
@@ -43,9 +99,8 @@ def test_vllm(mocker):
 	assert all(len(tokens[i]) == len(log_probs[i]) for i in range(len(tokens)))
 
 
-def test_vllm_chat_prompt(mocker):
-	mock_class = mocker.patch("vllm.LLM")
-	mock_class.return_value = LLM(model="mock-model")
+def test_vllm_chat_prompt(monkeypatch):
+	install_fake_vllm(monkeypatch)
 
 	previous_result = pd.DataFrame(
 		{"prompts": chat_prompts, "qid": ["id-1", "id-2", "id-3"]}
@@ -66,9 +121,8 @@ def test_vllm_chat_prompt(mocker):
 	assert all(len(tokens[i]) == len(log_probs[i]) for i in range(len(tokens)))
 
 
-def test_vllm_chat_prompt_think(mocker):
-	mock_class = mocker.patch("vllm.LLM")
-	mock_class.return_value = LLM(model="mock-model")
+def test_vllm_chat_prompt_think(monkeypatch):
+	install_fake_vllm(monkeypatch)
 
 	previous_result = pd.DataFrame(
 		{"prompts": chat_prompts, "qid": ["id-1", "id-2", "id-3"]}

autorag-0.3.26/tests/autorag/test_optional_extras.py

@@ -0,0 +1,49 @@
+import re
+from pathlib import Path
+
+LEGACY_ROOT = Path(__file__).resolve().parents[2]
+REPO_ROOT = Path(__file__).resolve().parents[3]
+
+
+def _extra_items(name: str) -> list[str]:
+	text = (LEGACY_ROOT / "pyproject.toml").read_text()
+	match = re.search(rf"^{name} = \[", text, re.M)
+	assert match is not None, f"missing optional extra {name}"
+	index = match.end()
+	depth = 1
+	cursor = index
+	while cursor < len(text) and depth:
+		char = text[cursor]
+		if char == "[":
+			depth += 1
+		elif char == "]":
+			depth -= 1
+		cursor += 1
+	body = text[index : cursor - 1]
+	return [
+		item.strip().strip("\"'")
+		for item in body.split(",")
+		if item.strip()
+	]
+
+
+def test_gpu_extra_does_not_include_vllm():
+	items = _extra_items("gpu")
+	assert not any(item.startswith("vllm") for item in items)
+
+
+def test_vllm_is_its_own_optional_extra():
+	assert any(item.startswith("vllm") for item in _extra_items("vllm"))
+	assert any("vllm" in item for item in _extra_items("all"))
+
+
+def test_legacy_unit_test_job_skips_vllm_install_and_bounds_runtime():
+	text = (REPO_ROOT / ".github/workflows/test.yml").read_text()
+	assert re.search(r"timeout-minutes:\s*[1-9]\d*", text)
+	sync_lines = [
+		line.strip() for line in text.splitlines() if "uv sync" in line
+	]
+	assert sync_lines, "legacy-test must install with uv sync"
+	assert all("--all-extras" not in line for line in sync_lines)
+	assert all("--extra vllm" not in line for line in sync_lines)
+	assert any("--extra gpu" in line for line in sync_lines)

{autorag-0.3.25 → autorag-0.3.26}/tests/resources/result_project/resources/chroma/6595d304-5270-4d9f-a267-c191366804ce/data_level0.bin

@@ -1,4 +1,4 @@
 [diffend] Oversized file quarantined before diffing.
-name: autorag-0.3.25/tests/resources/result_project/resources/chroma/6595d304-5270-4d9f-a267-c191366804ce/data_level0.bin
+name: autorag-0.3.26/tests/resources/result_project/resources/chroma/6595d304-5270-4d9f-a267-c191366804ce/data_level0.bin
 size: 62840000 bytes
 sha256: bb76ae0f6ca830a9a048b0cf53962a78c88c5c7fcda63fc846077d3456eb3890