{"nbformat":4,"nbformat_minor":0,"metadata":{"colab":{"provenance":[],"gpuType":"T4","authorship_tag":"ABX9TyPbfqCzoNG3OtcOe1gWk0cQ"},"kernelspec":{"name":"python3","display_name":"Python 3"},"language_info":{"name":"python"},"accelerator":"GPU"},"cells":[{"cell_type":"code","execution_count":null,"metadata":{"cellView":"form","id":"UT4uX7VzRvqQ"},"outputs":[],"source":["# ============================================================\n","# SANKAKU DATASET SEARCH + QWEN 3.5 CAPTIONING\n","# Google Colab T4 | Batched inference (20 images)\n","# ============================================================\n","# - Randomized Parquet traversal (seed-controlled)\n","# - Tag / rating / favourites filters\n","# - Image retrieval: deepghs/sankaku-webp-4Mpixel\n","# - Captioning: Qwen3.5-4B (4-bit), batches of 20\n","# - ZIP / CSV / JSONL export (same formats as before)\n","# - Optional local ZIP processed first, then Sankaku fills to IMAGE_COUNT\n","# - Optional word/symbol filter applied to final captions\n","# ============================================================\n","\n","# ---- installs ----\n","!pip -q install -U cheesechaser pyarrow pandas \"pillow<12.0\" \\\n"," \"transformers>=5.10.1\" accelerate bitsandbytes py7zr cryptography \\\n"," huggingface_hub torchvision safetensors sentencepiece packaging tqdm\n","!pip -q uninstall -y torchao\n","\n","# ============================================================\n","# COLAB PARAMETERS\n","# ============================================================\n","\n","SEARCH_QUERY = \"-ai-created \" #@param {type:\"string\"}\n","ADDITIONAL_WORDS = \"\" #@param {type:\"string\"}\n","\n","LOCAL_ZIP_PATH = \"\" #@param {type:\"string\"}\n","CAPTION_FILTER = \" \" #@param {type:\"string\"}\n","PROCESS_LOCAL_ZIP = True #@param {type:\"boolean\"}\n","\n","IMAGE_COUNT = 60 #@param {type:\"slider\", min:0, max:2000, step:30}\n","MIN_FAVOURITES = 0 #@param {type:\"slider\", min:0, max:1000, step:1}\n","\n","RANDOM_SEARCH_START = True #@param {type:\"boolean\"}\n","SEARCH_SEED_INPUT = \"ggvbbvbhh6\" #@param {type:\"string\"}\n","SEARCH_SKIP_AHEAD = 5 #@param {type:\"slider\", min:0, max:20, step:1}\n","\n","ENABLE_QWEN_CAPTIONER = True #@param {type:\"boolean\"}\n","\n","QWEN_MODEL_ID = \"DreamFast/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark\"\n","QWEN_BASE_MODEL_ID = \"Qwen/Qwen3.5-4B\"\n","\n","QWEN_MAX_NEW_TOKENS = 2048 #@param {type:\"slider\", min:64, max:2048, step:64}\n","QWEN_TEMPERATURE = 0.3 #@param {type:\"slider\", min:0.0, max:1.0, step:0.05}\n","QWEN_TOP_P = 0.90\n","QWEN_REPETITION_PENALTY = 1.05\n","QWEN_IMAGE_MAX_PIXELS = 262144\n","CAPTION_BATCH_SIZE = 30 # T4 batched captioning\n","\n","LORA_TEMPLATE_OUTPUT = True #@param {type:\"boolean\"}\n","KEEP_QWEN_IN_VRAM = True #@param {type:\"boolean\"}\n","\n","INCLUDE_SAFE = True #@param {type:\"boolean\"}\n","INCLUDE_QUESTIONABLE = True #@param {type:\"boolean\"}\n","INCLUDE_EXPLICIT = True #@param {type:\"boolean\"}\n","\n","# ============================================================\n","# SEARCH / EXPORT CONFIG\n","# ============================================================\n","\n","REPO_ID = \"deepghs/sankaku_full\"\n","WEBP_REPO_ID = \"deepghs/sankaku-webp-4Mpixel\"\n","\n","TEST_IMAGE_COUNT = int(IMAGE_COUNT)\n","TEST_MIN_FAVOURITES = int(MIN_FAVOURITES)\n","MAX_ROWS_TO_EXAMINE = 10_000_000\n","BATCH_SIZE = 5000 # parquet read batch\n","\n","SEED_FILE = \"/content/sankaku_search_seed.txt\"\n","GRID_MIN_WIDTH = 220\n","IMAGE_HEIGHT = 260\n","THUMBNAIL_MAX_SIZE = (1000, 1000)\n","THUMBNAIL_JPEG_QUALITY = 85\n","EXPORT_IMAGE_MAX_SIZE = (1024, 1024)\n","EXPORT_JPEG_QUALITY = 95\n","\n","OUTPUT_FILE = \"/content/sankaku_search_results.csv\"\n","JSONL_FILE = \"/content/sankaku_lora_dataset.jsonl\"\n","ZIP_FILE = (\n"," \"/content/sankaku_lora_training_template.zip\"\n"," if LORA_TEMPLATE_OUTPUT\n"," else \"/content/sankaku_image_text_pairs.zip\"\n",")\n","\n","INSTRUCTION_TEMPLATE = (\n"," \"Describe this image. Begin by enclosing the thinking in brackets. \"\n"," \"Then write Final output: followed by the response. The response text should \"\n"," \"be without newline, as a single block of text, 300-400 words in length. \"\n"," \"The output response should include these words: {tags} . \"\n"," \"Terminate the response by writing the word count. You may write Validation: \"\n"," \"afterwards if there is further thinking, which should be enclosed in \"\n"," \"brackets as well.\"\n",")\n","\n","# ============================================================\n","# IMPORTS\n","# ============================================================\n","\n","import os, sys, gc, io, re, html, json, base64, random, secrets, shutil\n","import zipfile, tempfile, warnings, traceback\n","from pathlib import Path\n","\n","import pyarrow.parquet as pq\n","import pandas as pd\n","import numpy as np\n","from PIL import Image, ImageOps, ImageFile\n","from tqdm.auto import tqdm\n","\n","from huggingface_hub import hf_hub_download, HfApi\n","from google.colab import userdata, output, files\n","from IPython.display import display, HTML\n","from cheesechaser.datapool.sankaku import SankakuWebpDataPool\n","\n","ImageFile.LOAD_TRUNCATED_IMAGES = True\n","warnings.filterwarnings(\"ignore\")\n","\n","# ============================================================\n","# SEED\n","# ============================================================\n","\n","def load_search_seed():\n"," if RANDOM_SEARCH_START:\n"," seed_text = secrets.token_hex(32)\n"," with open(SEED_FILE, \"w\", encoding=\"utf-8\") as f:\n"," f.write(seed_text)\n"," print(\"Random search enabled. Fresh seed written to\", SEED_FILE)\n"," else:\n"," seed_text = str(SEARCH_SEED_INPUT)\n"," print(\"Using manual seed:\", repr(seed_text))\n"," return seed_text\n","\n","ACTIVE_SEARCH_SEED = load_search_seed()\n","search_rng = random.Random(ACTIVE_SEARCH_SEED)\n","\n","# ============================================================\n","# AUTH\n","# ============================================================\n","\n","HF_TOKEN = userdata.get(\"HF_TOKEN\")\n","if not HF_TOKEN:\n"," raise RuntimeError(\"HF_TOKEN not found in Colab Secrets.\")\n","os.environ[\"HF_TOKEN\"] = HF_TOKEN\n","os.environ[\"HF_HUB_ENABLE_HF_TRANSFER\"] = \"1\"\n","\n","# ============================================================\n","# DISCOVER PARQUETS\n","# ============================================================\n","\n","print(\"\\n\" + \"=\" * 70)\n","print(\"DISCOVERING PARQUET FILES\")\n","print(\"=\" * 70)\n","\n","api = HfApi(token=HF_TOKEN)\n","repository_files = api.list_repo_files(repo_id=REPO_ID, repo_type=\"dataset\")\n","IMAGE_FILES = sorted(\n"," f for f in repository_files\n"," if f.startswith(\"tables/images-\") and f.lower().endswith(\".parquet\")\n",")\n","if not IMAGE_FILES:\n"," raise RuntimeError(f\"No tables/images-*.parquet files in {REPO_ID}.\")\n","\n","search_rng.shuffle(IMAGE_FILES)\n","print(\"Discovered Parquet files:\", len(IMAGE_FILES))\n","print(\"First file:\", IMAGE_FILES[0])\n","\n","# ============================================================\n","# QUERY PARSING\n","# ============================================================\n","\n","def split_search_words(text):\n"," if not text:\n"," return []\n"," return [w.strip().lower() for w in re.split(r\"[\\s+,\\.]+\", str(text)) if w.strip()]\n","\n","def parse_search_query(query):\n"," required, excluded = [], []\n"," query = str(query).lower()\n"," for match in re.finditer(r\"[^\\s+,\\.]+\", query):\n"," term = match.group(0).strip()\n"," if not term:\n"," continue\n"," starts_with_op = term.startswith(\"-\") and (\n"," match.start() == 0 or query[match.start() - 1].isspace()\n"," )\n"," if starts_with_op:\n"," word = term[1:].strip()\n"," if word:\n"," excluded.append(word)\n"," else:\n"," required.append(term)\n"," return required, excluded\n","\n","REQUIRED_TEST_TAGS, EXCLUDED_TEST_TAGS = parse_search_query(SEARCH_QUERY)\n","ADDITIONAL_TEST_TAGS = split_search_words(ADDITIONAL_WORDS)\n","if not REQUIRED_TEST_TAGS and not EXCLUDED_TEST_TAGS and not ADDITIONAL_TEST_TAGS:\n"," raise ValueError(\"Enter at least one main search term or additional word.\")\n","\n","ALLOWED_RATINGS = {\n"," r for r, en in ((\"s\", INCLUDE_SAFE), (\"q\", INCLUDE_QUESTIONABLE), (\"e\", INCLUDE_EXPLICIT)) if en\n","}\n","if not ALLOWED_RATINGS:\n"," raise ValueError(\"Enable at least one rating filter.\")\n","\n","# Caption filter tokens (case-insensitive whole-word / symbol removal)\n","FILTER_TOKENS = [t for t in re.split(r\"[\\s,]+\", str(CAPTION_FILTER or \"\").strip()) if t]\n","print(\"=\" * 70)\n","print(\"SANKAKU IMAGE SEARCH + QWEN 3.5 CAPTIONING\")\n","print(\"=\" * 70)\n","print(\"Required:\", REQUIRED_TEST_TAGS)\n","print(\"Excluded:\", EXCLUDED_TEST_TAGS)\n","print(\"Additional (OR):\", ADDITIONAL_TEST_TAGS)\n","print(\"Images requested:\", TEST_IMAGE_COUNT)\n","print(\"Min favourites:\", TEST_MIN_FAVOURITES)\n","print(\"Qwen captioner:\", ENABLE_QWEN_CAPTIONER)\n","print(\"Caption batch size:\", CAPTION_BATCH_SIZE)\n","print(\"Local ZIP processing:\", PROCESS_LOCAL_ZIP, \"| path:\", LOCAL_ZIP_PATH or \"(none)\")\n","print(\"Caption filter tokens:\", FILTER_TOKENS or \"(none)\")\n","\n","# ============================================================\n","# TAG / MATCH HELPERS\n","# ============================================================\n","\n","def tags_to_search_text(tags):\n"," if tags is None:\n"," return \"\"\n"," if isinstance(tags, (list, tuple, set)):\n"," return \" \".join(str(t) for t in tags if t is not None).lower()\n"," return str(tags).lower()\n","\n","def tags_to_display_list(tags):\n"," if tags is None:\n"," return []\n"," if isinstance(tags, (list, tuple, set)):\n"," return [str(t) for t in tags if t is not None]\n"," return str(tags).split()\n","\n","def matches_tag_conditions(tags):\n"," tag_text = tags_to_search_text(tags)\n"," if not tag_text:\n"," return False\n"," if not all(t in tag_text for t in REQUIRED_TEST_TAGS):\n"," return False\n"," if any(t in tag_text for t in EXCLUDED_TEST_TAGS):\n"," return False\n"," if ADDITIONAL_TEST_TAGS and not any(t in tag_text for t in ADDITIONAL_TEST_TAGS):\n"," return False\n"," return True\n","\n","def normalize_fav_count(value):\n"," if value is None:\n"," return 0\n"," try:\n"," if pd.isna(value):\n"," return 0\n"," return max(0, int(value))\n"," except (TypeError, ValueError, OverflowError):\n"," return 0\n","\n","def normalize_rating(value):\n"," if value is None:\n"," return \"\"\n"," try:\n"," if pd.isna(value):\n"," return \"\"\n"," except (TypeError, ValueError):\n"," pass\n"," return str(value).strip().lower()\n","\n","def tags_to_prompt_string(tags):\n"," if isinstance(tags, str):\n"," tags = tags.replace(\",\", \" \").split()\n"," tag_list = [str(t).strip() for t in (tags or []) if str(t).strip()]\n"," return \" \".join(tag_list)\n","\n","def build_caption_prompt(tags):\n"," tag_str = tags_to_prompt_string(tags)\n"," return INSTRUCTION_TEMPLATE.format(tags=tag_str)\n","\n","def normalize_caption_text(text):\n"," if text is None:\n"," return \"\"\n"," return re.sub(r\"\\s+\", \" \", str(text)).strip()\n","\n","def apply_caption_filter(text):\n"," \"\"\"Remove configured words/symbols from caption text.\"\"\"\n"," if not text or not FILTER_TOKENS:\n"," return normalize_caption_text(text)\n"," out = str(text)\n"," for tok in FILTER_TOKENS:\n"," # Escape for regex; treat as literal token (word boundary when alphanumeric)\n"," escaped = re.escape(tok)\n"," if re.match(r\"^[\\w'-]+$\", tok):\n"," pattern = rf\"\\b{escaped}\\b\"\n"," else:\n"," pattern = escaped\n"," out = re.sub(pattern, \" \", out, flags=re.IGNORECASE)\n"," return normalize_caption_text(out)\n","\n","# ============================================================\n","# LOCAL ZIP PRE-PROCESSING (runs before Sankaku search)\n","# ============================================================\n","\n","test_results = [] # will hold both local-zip and Sankaku items\n","local_zip_count = 0\n","image_data_uris = {}\n","downloaded_bytes = 0\n","failed_images = []\n","caption_failures = []\n","EXPORT_DIR = tempfile.mkdtemp(prefix=\"sankaku_export_\")\n","export_image_count = 0\n","pending_caption_jobs = [] # list of dicts: item, image_id, pil\n","\n","def find_image_file(directory):\n"," candidates = []\n"," for root, _, files_in_dir in os.walk(directory):\n"," for filename in files_in_dir:\n"," if filename.endswith(\"_metainfo.json\"):\n"," continue\n"," candidates.append(os.path.join(root, filename))\n"," if not candidates:\n"," raise RuntimeError(\"SankakuWebpDataPool returned no files.\")\n"," preferred = [\".webp\", \".jpg\", \".jpeg\", \".png\", \".gif\", \".bmp\", \".tif\", \".tiff\"]\n"," def prio(path):\n"," ext = os.path.splitext(path)[1].lower()\n"," try:\n"," return preferred.index(ext)\n"," except ValueError:\n"," return len(preferred)\n"," candidates.sort(key=lambda p: (prio(p), p))\n"," for fp in candidates:\n"," try:\n"," with Image.open(fp) as im:\n"," im.verify()\n"," return fp\n"," except Exception:\n"," continue\n"," raise RuntimeError(\"No decodable image from WebP DataPool.\")\n","\n","def make_thumbnail_data_uri(image_path):\n"," with Image.open(image_path) as source:\n"," source.load()\n"," source = ImageOps.exif_transpose(source)\n"," thumb = source.convert(\"RGB\")\n"," thumb.thumbnail(THUMBNAIL_MAX_SIZE, Image.Resampling.LANCZOS)\n"," buf = io.BytesIO()\n"," thumb.save(buf, format=\"JPEG\", quality=THUMBNAIL_JPEG_QUALITY, optimize=True)\n"," return \"data:image/jpeg;base64,\" + base64.b64encode(buf.getvalue()).decode(\"ascii\")\n","\n","def prepare_export_image(image_path):\n"," with Image.open(image_path) as source:\n"," source.load()\n"," source = ImageOps.exif_transpose(source)\n"," export_image = source.convert(\"RGB\")\n"," export_image.thumbnail(EXPORT_IMAGE_MAX_SIZE, Image.Resampling.LANCZOS)\n"," buf = io.BytesIO()\n"," export_image.save(\n"," buf, format=\"JPEG\", quality=EXPORT_JPEG_QUALITY, optimize=True, progressive=True\n"," )\n"," return buf.getvalue()\n","\n","def make_export_text(item):\n"," natural = normalize_caption_text(item.get(\"qwen_caption\", \"\"))\n"," tags = item.get(\"tags\", [])\n"," if isinstance(tags, str):\n"," tags = tags.replace(\",\", \" \").split()\n"," tags_text = \" \".join(str(t).strip() for t in tags if str(t).strip())\n"," if natural and tags_text:\n"," return f\"{natural} {tags_text}\"\n"," return natural or tags_text\n","\n","if PROCESS_LOCAL_ZIP and LOCAL_ZIP_PATH and os.path.isfile(LOCAL_ZIP_PATH):\n"," print(\"\\n\" + \"=\" * 70)\n"," print(\"PROCESSING LOCAL ZIP FIRST\")\n"," print(\"=\" * 70)\n"," print(\"ZIP:\", LOCAL_ZIP_PATH)\n","\n"," with tempfile.TemporaryDirectory(prefix=\"local_zip_\") as tmp_extract:\n"," try:\n"," with zipfile.ZipFile(LOCAL_ZIP_PATH, \"r\") as zf:\n"," zf.extractall(tmp_extract)\n"," except Exception as e:\n"," print(\"Failed to extract local ZIP:\", e)\n"," else:\n"," # Collect image files recursively\n"," image_exts = {\".jpg\", \".jpeg\", \".png\", \".webp\", \".bmp\", \".gif\", \".tif\", \".tiff\"}\n"," local_images = []\n"," for root, _, files_in_dir in os.walk(tmp_extract):\n"," for fn in files_in_dir:\n"," if os.path.splitext(fn)[1].lower() in image_exts:\n"," local_images.append(os.path.join(root, fn))\n"," local_images.sort()\n","\n"," # Limit to remaining slots (we may still need Sankaku later)\n"," max_from_zip = TEST_IMAGE_COUNT\n"," local_images = local_images[:max_from_zip]\n"," print(f\"Found {len(local_images)} image(s) in ZIP (using up to {max_from_zip})\")\n","\n"," for idx, img_path in enumerate(local_images, 1):\n"," # Synthetic ID so it never collides with real Sankaku IDs\n"," image_id = f\"local_{idx:05d}\"\n"," print(f\"\\n[local {idx}/{len(local_images)}] {os.path.basename(img_path)}\")\n"," try:\n"," original_size = os.path.getsize(img_path)\n"," downloaded_bytes += original_size\n"," with Image.open(img_path) as im:\n"," print(f\" {original_size/1024:.1f} KB | {im.format} | {im.size[0]}×{im.size[1]}\")\n","\n"," image_data_uris[image_id] = make_thumbnail_data_uri(img_path)\n","\n"," # Minimal result dict (no real tags / rating / favs)\n"," item = {\n"," \"id\": image_id,\n"," \"tags\": [],\n"," \"rating\": \"\",\n"," \"width\": None,\n"," \"height\": None,\n"," \"created_at\": None,\n"," \"fav_count\": 0,\n"," \"author_name\": \"local_zip\",\n"," \"qwen_caption\": \"\",\n"," \"input_text\": build_caption_prompt([]),\n"," \"output_text\": \"\",\n"," \"image\": \"\",\n"," \"source\": \"local_zip\",\n"," }\n","\n"," # export jpg\n"," image_export_path = os.path.join(EXPORT_DIR, f\"{image_id}.jpg\")\n"," try:\n"," export_bytes = prepare_export_image(img_path)\n"," with open(image_export_path, \"wb\") as f:\n"," f.write(export_bytes)\n"," item[\"image\"] = f\"{image_id}.jpg\"\n"," export_image_count += 1\n"," except Exception as e:\n"," print(\"IMAGE EXPORT FAILED:\", e)\n"," item[\"image\"] = \"\"\n"," if os.path.exists(image_export_path):\n"," os.remove(image_export_path)\n","\n"," if ENABLE_QWEN_CAPTIONER and item.get(\"image\"):\n"," with Image.open(img_path) as source:\n"," source.load()\n"," caption_image = ImageOps.exif_transpose(source).convert(\"RGB\")\n"," pending_caption_jobs.append({\n"," \"item\": item,\n"," \"image_id\": image_id,\n"," \"pil\": caption_image.copy(),\n"," })\n"," caption_image.close()\n"," else:\n"," item[\"qwen_caption\"] = \"\"\n"," item[\"output_text\"] = \"\"\n","\n"," test_results.append(item)\n"," local_zip_count += 1\n","\n"," except Exception as exc:\n"," failed_images.append(image_id)\n"," image_data_uris.pop(image_id, None)\n"," print(f\"LOCAL IMAGE FAILED: {type(exc).__name__}: {exc}\")\n"," gc.collect()\n","\n"," print(f\"\\nLocal ZIP contributed {local_zip_count} image(s).\")\n","else:\n"," if PROCESS_LOCAL_ZIP and LOCAL_ZIP_PATH:\n"," print(f\"\\nPROCESS_LOCAL_ZIP=True but path not found: {LOCAL_ZIP_PATH}\")\n"," else:\n"," print(\"\\nLocal ZIP processing skipped.\")\n","\n","# How many more images Sankaku still needs to fetch\n","remaining_needed = max(0, TEST_IMAGE_COUNT - local_zip_count)\n","print(f\"Sankaku will fetch up to {remaining_needed} additional image(s) (total target {TEST_IMAGE_COUNT}).\")\n","\n","# ============================================================\n","# SEARCH PARQUET\n","# ============================================================\n","\n","print(\"\\n\" + \"=\" * 70)\n","print(\"SEARCHING PARQUET FILES\")\n","print(\"=\" * 70)\n","\n","rows_examined = rows_skipped = batches_examined = 0\n","tag_matches = rating_rejected = favourites_rejected = 0\n","search_finished = remaining_needed == 0\n","search_stop_reason = \"No additional images needed (local ZIP already filled quota)\" if search_finished else \"Search not started\"\n","skip_remaining = 0\n","initial_offset_remaining = 0\n","initial_offset_set = False\n","row_group_order_metadata = []\n","\n","for image_filename in IMAGE_FILES:\n"," if search_finished:\n"," break\n"," print(\"\\nOpening:\", image_filename)\n"," image_file = hf_hub_download(\n"," repo_id=REPO_ID, repo_type=\"dataset\", filename=image_filename, token=HF_TOKEN\n"," )\n"," pf = pq.ParquetFile(image_file)\n"," available_columns = set(pf.schema_arrow.names)\n"," wanted = [\"id\", \"tags\", \"rating\", \"width\", \"height\", \"created_at\", \"fav_count\", \"author_name\"]\n"," columns = [c for c in wanted if c in available_columns]\n"," missing = {\"id\", \"tags\", \"fav_count\", \"rating\"} - set(columns)\n"," if missing:\n"," raise RuntimeError(f\"{image_filename} missing columns: {sorted(missing)}\")\n","\n"," print(\"Rows:\", f\"{pf.metadata.num_rows:,}\", \"| row groups:\", pf.num_row_groups)\n"," row_groups = list(range(pf.num_row_groups))\n"," search_rng.shuffle(row_groups)\n"," row_group_order_metadata.append({\"file\": image_filename, \"row_groups\": list(row_groups)})\n","\n"," if not initial_offset_set and row_groups:\n"," first_group_rows = pf.metadata.row_group(row_groups[0]).num_rows\n"," initial_offset_remaining = search_rng.randrange(first_group_rows) if first_group_rows > 0 else 0\n"," initial_offset_set = True\n"," print(\"Random starting row offset:\", initial_offset_remaining)\n","\n"," for rg_i, row_group in enumerate(row_groups, 1):\n"," if search_finished:\n"," break\n"," for batch in pf.iter_batches(\n"," batch_size=BATCH_SIZE, row_groups=[row_group], columns=columns, use_threads=True\n"," ):\n"," batches_examined += 1\n"," data = batch.to_pydict()\n"," for i, image_tags in enumerate(data[\"tags\"]):\n"," if initial_offset_remaining > 0:\n"," initial_offset_remaining -= 1\n"," rows_skipped += 1\n"," continue\n"," if skip_remaining > 0:\n"," skip_remaining -= 1\n"," rows_skipped += 1\n"," continue\n"," if rows_examined >= MAX_ROWS_TO_EXAMINE:\n"," search_finished = True\n"," search_stop_reason = \"Maximum examined-row limit reached\"\n"," break\n"," rows_examined += 1\n","\n"," if matches_tag_conditions(image_tags):\n"," tag_matches += 1\n"," image_rating = normalize_rating(data[\"rating\"][i])\n"," if image_rating not in ALLOWED_RATINGS:\n"," rating_rejected += 1\n"," else:\n"," fav_count = normalize_fav_count(data[\"fav_count\"][i])\n"," if fav_count < TEST_MIN_FAVOURITES:\n"," favourites_rejected += 1\n"," else:\n"," image_id_value = data[\"id\"][i]\n"," if image_id_value is not None:\n"," result = {}\n"," for column in columns:\n"," value = data[column][i]\n"," if column == \"tags\":\n"," result[column] = tags_to_display_list(image_tags)\n"," elif column == \"fav_count\":\n"," result[column] = fav_count\n"," else:\n"," result[column] = value\n"," result[\"qwen_caption\"] = \"\"\n"," result[\"input_text\"] = \"\"\n"," result[\"output_text\"] = \"\"\n"," result[\"image\"] = \"\"\n"," result[\"source\"] = \"sankaku\"\n"," test_results.append(result)\n"," print(\n"," f\" MATCH {len(test_results)}/{TEST_IMAGE_COUNT}: \"\n"," f\"ID {result['id']} ({fav_count} favs)\"\n"," )\n"," skip_remaining = int(SEARCH_SKIP_AHEAD)\n"," # Count only Sankaku items toward remaining quota\n"," sankaku_count = sum(1 for r in test_results if r.get(\"source\") == \"sankaku\")\n"," if sankaku_count >= remaining_needed:\n"," search_finished = True\n"," search_stop_reason = \"Requested image count reached\"\n"," break\n"," if rows_examined >= MAX_ROWS_TO_EXAMINE:\n"," search_finished = True\n"," search_stop_reason = \"Maximum examined-row limit reached\"\n"," break\n"," del data\n"," if search_finished:\n"," break\n"," if search_finished:\n"," break\n"," del pf\n"," gc.collect()\n"," if search_finished:\n"," break\n","\n","print(\"\\nPARQUET SEARCH SUMMARY\")\n","print(\"Tag matches:\", tag_matches)\n","print(\"Total items (local + Sankaku):\", len(test_results))\n","print(\" - from local ZIP:\", local_zip_count)\n","print(\" - from Sankaku:\", sum(1 for r in test_results if r.get(\"source\") == \"sankaku\"))\n","print(\"Rows examined:\", f\"{rows_examined:,}\")\n","print(\"Stop reason:\", search_stop_reason)\n","if not test_results:\n"," raise RuntimeError(\"No images matched search conditions and no usable local ZIP images.\")\n","\n","# ============================================================\n","# QWEN MODEL STATE / LOAD / UNLOAD\n","# ============================================================\n","\n","if \"qwen_model\" not in globals():\n"," qwen_model = None\n","if \"qwen_processor\" not in globals():\n"," qwen_processor = None\n","if \"qwen_tokenizer\" not in globals():\n"," qwen_tokenizer = None\n","if \"qwen_loaded_model_id\" not in globals():\n"," qwen_loaded_model_id = None\n","\n","def unload_qwen():\n"," global qwen_model, qwen_processor, qwen_tokenizer, qwen_loaded_model_id\n"," print(\"\\nUNLOADING QWEN\")\n"," qwen_model = qwen_processor = qwen_tokenizer = None\n"," qwen_loaded_model_id = None\n"," gc.collect()\n"," try:\n"," import torch\n"," if torch.cuda.is_available():\n"," torch.cuda.synchronize()\n"," torch.cuda.empty_cache()\n"," print(f\"GPU allocated: {torch.cuda.memory_allocated() / 1024**3:.2f} GB\")\n"," except Exception as e:\n"," print(\"GPU cleanup warning:\", e)\n","\n","def load_qwen():\n"," global qwen_model, qwen_processor, qwen_tokenizer, qwen_loaded_model_id\n"," import torch\n"," from transformers import (\n"," AutoProcessor, AutoTokenizer, AutoModelForMultimodalLM, BitsAndBytesConfig\n"," )\n","\n"," print(\"\\n\" + \"=\" * 70)\n"," print(\"QWEN 3.5 CAPTIONER\")\n"," print(\"=\" * 70)\n","\n"," if not torch.cuda.is_available():\n"," raise RuntimeError(\"CUDA GPU required. Enable T4 runtime.\")\n","\n"," requested = f\"{QWEN_MODEL_ID}\"\n"," if (\n"," qwen_model is not None\n"," and qwen_processor is not None\n"," and qwen_loaded_model_id == requested\n"," ):\n"," print(\"Reusing Qwen already in VRAM:\", qwen_loaded_model_id)\n"," return\n","\n"," if qwen_model is not None:\n"," unload_qwen()\n","\n"," print(\"Loading:\", QWEN_MODEL_ID)\n"," tokenizer = AutoTokenizer.from_pretrained(\n"," QWEN_BASE_MODEL_ID, token=HF_TOKEN, trust_remote_code=True\n"," )\n"," processor = AutoProcessor.from_pretrained(\n"," QWEN_MODEL_ID, token=HF_TOKEN, trust_remote_code=True\n"," )\n"," try:\n"," tokenizer.padding_side = \"left\"\n"," processor.tokenizer = tokenizer\n"," processor.chat_template = tokenizer.chat_template\n"," except Exception:\n"," pass\n"," try:\n"," processor.image_processor.max_pixels = QWEN_IMAGE_MAX_PIXELS\n"," except Exception:\n"," pass\n","\n"," qcfg = BitsAndBytesConfig(\n"," load_in_4bit=True,\n"," bnb_4bit_quant_type=\"nf4\",\n"," bnb_4bit_use_double_quant=True,\n"," bnb_4bit_compute_dtype=torch.float16,\n"," )\n"," model = AutoModelForMultimodalLM.from_pretrained(\n"," QWEN_MODEL_ID,\n"," dtype=torch.float16,\n"," quantization_config=qcfg,\n"," device_map={\"\": 0},\n"," low_cpu_mem_usage=True,\n"," token=HF_TOKEN,\n"," attn_implementation=\"sdpa\",\n"," trust_remote_code=True,\n"," )\n"," model.eval()\n"," qwen_model = model\n"," qwen_processor = processor\n"," qwen_tokenizer = tokenizer\n"," qwen_loaded_model_id = requested\n"," print(\"Qwen loaded.\")\n"," print(f\"GPU allocated: {torch.cuda.memory_allocated() / 1024**3:.2f} GB\")\n","\n","def split_response(raw):\n"," raw = str(raw or \"\").replace(\"<|im_start|>\", \"\\n\").replace(\"<|im_end|>\", \"\\n\")\n"," raw = raw.replace(\"<|fim_suffix|>\", \"\\n\")\n"," raw = re.sub(r\"<\\|(?:assistant|user|system)\\|>\", \"\\n\", raw)\n"," blocks = re.findall(r\"(.*?)\", raw, flags=re.I | re.S)\n"," thinking = blocks[0].strip() if blocks else \"\"\n","\n"," marker = re.search(r\"Final\\s+output\\s*:\\s*\", raw, flags=re.I)\n"," if marker:\n"," final = raw[marker.end():]\n"," elif \"\" in raw.lower():\n"," final = raw[raw.lower().rfind(\"\") + len(\"\"):]\n"," else:\n"," final = raw\n","\n"," final = re.sub(r\"<\\|[^>]+\\|>\", \" \", final)\n"," addendum = \"\"\n"," validation = re.split(r\"\\bValidation\\s*:\\s*\", final, maxsplit=1, flags=re.I)\n"," if len(validation) == 2:\n"," final, addendum = validation[0], validation[1].strip()\n"," if len(blocks) > 1:\n"," addendum = (addendum + \"\\n\" + \"\\n\".join(blocks[1:])).strip()\n"," final = re.sub(r\".*?\", \" \", final, flags=re.I | re.S)\n"," final = re.sub(r\"^\\s*Final\\s+output\\s*:\\s*\", \"\", final, flags=re.I)\n"," final = re.sub(r\"\\s+\", \" \", final).strip()\n"," words = re.findall(r\"\\b[\\w'-]+\\b\", final)\n"," if words and not re.search(r\"\\b\\d+\\s+words?\\b\\s*$\", final, flags=re.I):\n"," final += f\" ({len(words)} words)\"\n"," return thinking, final, addendum\n","\n","def qwen_caption_batch(images, tags_list):\n"," \"\"\"Caption a list of PIL images (len <= CAPTION_BATCH_SIZE).\"\"\"\n"," import torch\n","\n"," if qwen_model is None or qwen_processor is None:\n"," raise RuntimeError(\"Qwen not loaded.\")\n","\n"," conversations = []\n"," prompts = []\n"," for image, tags in zip(images, tags_list):\n"," instruction = build_caption_prompt(tags)\n"," prompts.append(instruction)\n"," conversations.append([\n"," {\n"," \"role\": \"user\",\n"," \"content\": [\n"," {\"type\": \"image\", \"image\": image},\n"," {\"type\": \"text\", \"text\": instruction},\n"," ],\n"," }\n"," ])\n","\n"," inputs = qwen_processor.apply_chat_template(\n"," conversations,\n"," chat_template=qwen_tokenizer.chat_template,\n"," tokenize=True,\n"," add_generation_prompt=True,\n"," return_dict=True,\n"," return_tensors=\"pt\",\n"," padding=True,\n"," )\n"," if not hasattr(inputs, \"items\"):\n"," raise TypeError(f\"Expected dict-like inputs, got {type(inputs)}\")\n","\n"," device = torch.device(\"cuda:0\")\n"," inputs = {k: (v.to(device) if hasattr(v, \"to\") else v) for k, v in inputs.items()}\n","\n"," kwargs = {\n"," \"max_new_tokens\": int(QWEN_MAX_NEW_TOKENS),\n"," \"repetition_penalty\": float(QWEN_REPETITION_PENALTY),\n"," \"do_sample\": QWEN_TEMPERATURE > 0,\n"," \"use_cache\": True,\n"," \"pad_token_id\": qwen_tokenizer.pad_token_id,\n"," }\n"," if QWEN_TEMPERATURE > 0:\n"," kwargs.update(temperature=float(QWEN_TEMPERATURE), top_p=float(QWEN_TOP_P))\n","\n"," with torch.inference_mode():\n"," output = qwen_model.generate(**inputs, **kwargs)\n","\n"," prompt_len = inputs[\"input_ids\"].shape[-1]\n"," generated = output[:, prompt_len:]\n"," raws = qwen_processor.batch_decode(\n"," generated, skip_special_tokens=False, clean_up_tokenization_spaces=False\n"," )\n"," results = []\n"," for raw, prompt in zip(raws, prompts):\n"," thinking, written, addendum = split_response(raw)\n"," # Apply user-configured filter to the final caption\n"," filtered = apply_caption_filter(written)\n"," results.append({\n"," \"input_text\": prompt,\n"," \"thinking\": thinking,\n"," \"caption\": filtered,\n"," \"addendum\": addendum,\n"," \"raw\": raw,\n"," })\n"," del inputs, output, generated\n"," torch.cuda.empty_cache()\n"," return results\n","\n","if ENABLE_QWEN_CAPTIONER:\n"," load_qwen()\n","else:\n"," print(\"\\nQwen captioning disabled.\")\n","\n","# ============================================================\n","# WEBP POOL + RETRIEVE IMAGES (Sankaku only)\n","# ============================================================\n","\n","print(\"\\n\" + \"=\" * 70)\n","print(\"FETCHING WEBP IMAGES (Sankaku)\")\n","print(\"=\" * 70)\n","\n","pool = SankakuWebpDataPool(hf_token=HF_TOKEN)\n","\n","# ---- retrieve Sankaku images ----\n","sankaku_items = [item for item in test_results if item.get(\"source\") == \"sankaku\"]\n","for number, item in enumerate(sankaku_items, 1):\n"," image_id = str(int(item[\"id\"]))\n"," print(f\"\\n[{number}/{len(sankaku_items)}] Retrieving WebP ID {image_id}\")\n"," try:\n"," with pool.mock_resource(int(image_id), None, silent=True) as (temp_dir, _info):\n"," image_path = find_image_file(temp_dir)\n"," original_size = os.path.getsize(image_path)\n"," downloaded_bytes += original_size\n"," with Image.open(image_path) as im:\n"," print(f\" {original_size/1024:.1f} KB | {im.format} | {im.size[0]}×{im.size[1]}\")\n","\n"," image_data_uris[image_id] = make_thumbnail_data_uri(image_path)\n"," item[\"input_text\"] = build_caption_prompt(item.get(\"tags\", []))\n","\n"," # export jpg\n"," image_export_path = os.path.join(EXPORT_DIR, f\"{image_id}.jpg\")\n"," try:\n"," export_bytes = prepare_export_image(image_path)\n"," with open(image_export_path, \"wb\") as f:\n"," f.write(export_bytes)\n"," item[\"image\"] = f\"{image_id}.jpg\"\n"," export_image_count += 1\n"," except Exception as e:\n"," print(\"IMAGE EXPORT FAILED:\", e)\n"," item[\"image\"] = \"\"\n"," if os.path.exists(image_export_path):\n"," os.remove(image_export_path)\n","\n"," if ENABLE_QWEN_CAPTIONER and item.get(\"image\"):\n"," with Image.open(image_path) as source:\n"," source.load()\n"," caption_image = ImageOps.exif_transpose(source).convert(\"RGB\")\n"," pending_caption_jobs.append({\n"," \"item\": item,\n"," \"image_id\": image_id,\n"," \"pil\": caption_image.copy(),\n"," })\n"," caption_image.close()\n"," else:\n"," item[\"qwen_caption\"] = \"\"\n"," item[\"output_text\"] = \"\"\n","\n"," except Exception as exc:\n"," failed_images.append(image_id)\n"," image_data_uris.pop(image_id, None)\n"," item[\"qwen_caption\"] = item[\"input_text\"] = item[\"output_text\"] = item[\"image\"] = \"\"\n"," print(f\"IMAGE RETRIEVAL FAILED: {type(exc).__name__}: {exc}\")\n"," gc.collect()\n","\n","# ---- batched Qwen captioning (covers both local-zip and Sankaku jobs) ----\n","if ENABLE_QWEN_CAPTIONER and pending_caption_jobs:\n"," print(\"\\n\" + \"=\" * 70)\n"," print(f\"BATCHED QWEN CAPTIONING ({CAPTION_BATCH_SIZE} images / batch)\")\n"," print(\"=\" * 70)\n"," total = len(pending_caption_jobs)\n"," for start in tqdm(range(0, total, CAPTION_BATCH_SIZE), desc=\"Caption batches\"):\n"," batch_jobs = pending_caption_jobs[start:start + CAPTION_BATCH_SIZE]\n"," images = [j[\"pil\"] for j in batch_jobs]\n"," tags_list = [j[\"item\"].get(\"tags\", []) for j in batch_jobs]\n"," try:\n"," results = qwen_caption_batch(images, tags_list)\n"," for job, res in zip(batch_jobs, results):\n"," item = job[\"item\"]\n"," item[\"input_text\"] = res[\"input_text\"]\n"," item[\"qwen_caption\"] = res[\"caption\"]\n"," item[\"output_text\"] = res[\"caption\"]\n"," if not res[\"caption\"].strip():\n"," caption_failures.append(job[\"image_id\"])\n"," print(f\" WARNING empty caption for ID {job['image_id']}\")\n"," else:\n"," print(f\" OK ID {job['image_id']}: {res['caption'][:120]}...\")\n"," except Exception as e:\n"," print(f\"BATCH FAILED: {e}\")\n"," traceback.print_exc()\n"," for job in batch_jobs:\n"," job[\"item\"][\"qwen_caption\"] = \"\"\n"," job[\"item\"][\"output_text\"] = \"\"\n"," caption_failures.append(job[\"image_id\"])\n"," try:\n"," import torch\n"," torch.cuda.empty_cache()\n"," except Exception:\n"," pass\n"," finally:\n"," for j in batch_jobs:\n"," try:\n"," j[\"pil\"].close()\n"," except Exception:\n"," pass\n"," gc.collect()\n"," try:\n"," import torch\n"," torch.cuda.empty_cache()\n"," except Exception:\n"," pass\n","\n","# ============================================================\n","# QWEN VRAM MANAGEMENT\n","# ============================================================\n","\n","print(\"\\n\" + \"=\" * 70)\n","print(\"QWEN MEMORY MANAGEMENT\")\n","print(\"=\" * 70)\n","if ENABLE_QWEN_CAPTIONER:\n"," if KEEP_QWEN_IN_VRAM:\n"," print(\"KEEP_QWEN_IN_VRAM=True — model stays loaded.\")\n"," try:\n"," import torch\n"," if torch.cuda.is_available():\n"," print(f\"GPU allocated: {torch.cuda.memory_allocated()/1024**3:.2f} GB\")\n"," except Exception:\n"," pass\n"," else:\n"," unload_qwen()\n","else:\n"," print(\"Captioning was disabled.\")\n","\n","# ============================================================\n","# JSONL / CSV / ZIP (same structure as original)\n","# ============================================================\n","\n","training_records = []\n","for item in test_results:\n"," image_filename = str(item.get(\"image\", \"\") or \"\").strip()\n"," input_text = str(item.get(\"input_text\", \"\") or \"\").strip()\n"," output_text = normalize_caption_text(item.get(\"output_text\", \"\"))\n"," if not image_filename or not input_text or not output_text:\n"," continue\n"," image_path = os.path.join(EXPORT_DIR, image_filename)\n"," if not os.path.isfile(image_path):\n"," continue\n"," training_records.append({\n"," \"image\": image_filename,\n"," \"input_text\": input_text,\n"," \"output_text\": output_text,\n"," })\n","\n","print(\"\\nLoRA JSONL examples:\", len(training_records))\n","print(\"Exported JPGs:\", export_image_count)\n","\n","with open(JSONL_FILE, \"w\", encoding=\"utf-8\") as f:\n"," for rec in training_records:\n"," f.write(json.dumps(rec, ensure_ascii=False) + \"\\n\")\n","print(\"JSONL saved:\", JSONL_FILE)\n","\n","row_group_order_text = \"\\n\".join(\n"," f\"{e['file']}: \" + \", \".join(map(str, e[\"row_groups\"]))\n"," for e in row_group_order_metadata\n",")\n","\n","if LORA_TEMPLATE_OUTPUT:\n"," zip_format_description = \"Numbered LoRA input/output triplets.\"\n"," zip_contents_description = \"Only numbered JPG/input/output triplets.\"\n"," zip_pairing_description = (\n"," \"N.jpg + N_input.txt + N_output.txt per image. \"\n"," \"Input = exact Qwen prompt. Output = Qwen caption (empty if failed).\"\n"," )\n","else:\n"," zip_format_description = \"Original Sankaku ID image-text pairs.\"\n"," zip_contents_description = \"ID.jpg/ID.txt + JSONL + search_metadata.txt\"\n"," zip_pairing_description = (\n"," \"TXT = Qwen caption + original tags (or tags only if no caption).\"\n"," )\n","\n","search_metadata = f\"\"\"\n","SANKAKU SEARCH + QWEN 3.5 CAPTION METADATA\n","==========================================\n","Metadata repo: {REPO_ID}\n","Image repo: {WEBP_REPO_ID}\n","Query: {SEARCH_QUERY}\n","Required: {REQUIRED_TEST_TAGS}\n","Excluded: {EXCLUDED_TEST_TAGS}\n","Additional: {ADDITIONAL_TEST_TAGS}\n","Min favourites: {TEST_MIN_FAVOURITES}\n","Images requested: {TEST_IMAGE_COUNT}\n","Local ZIP path: {LOCAL_ZIP_PATH or \"(none)\"}\n","Local ZIP processed: {PROCESS_LOCAL_ZIP}\n","Local ZIP images used: {local_zip_count}\n","Rows examined: {rows_examined}\n","Rows skipped: {rows_skipped}\n","Stop reason: {search_stop_reason}\n","Active seed: {ACTIVE_SEARCH_SEED!r}\n","Qwen model: {QWEN_MODEL_ID}\n","Qwen base: {QWEN_BASE_MODEL_ID}\n","Max new tokens: {QWEN_MAX_NEW_TOKENS}\n","Temperature: {QWEN_TEMPERATURE}\n","Caption batch size: {CAPTION_BATCH_SIZE}\n","Caption filter: {FILTER_TOKENS or \"(none)\"}\n","Exported JPGs: {export_image_count}\n","JSONL records: {len(training_records)}\n","Caption failures: {len(caption_failures)}\n","ZIP mode: {zip_format_description}\n","LORA_TEMPLATE_OUTPUT: {LORA_TEMPLATE_OUTPUT}\n","\"\"\".strip()\n","\n","print(\"\\n\" + \"=\" * 70)\n","print(\"CREATING ZIP\")\n","print(\"=\" * 70)\n","if os.path.exists(ZIP_FILE):\n"," os.remove(ZIP_FILE)\n","\n","complete_image_ids = []\n","for item in test_results:\n"," # Support both numeric Sankaku IDs and synthetic local_xxxxx IDs\n"," img_id = str(item[\"id\"])\n"," if os.path.isfile(os.path.join(EXPORT_DIR, f\"{img_id}.jpg\")):\n"," complete_image_ids.append(img_id)\n","complete_image_ids = list(dict.fromkeys(complete_image_ids))\n","export_image_count = len(complete_image_ids)\n","items_by_image_id = {str(item[\"id\"]): item for item in test_results}\n","\n","with zipfile.ZipFile(ZIP_FILE, \"w\", compression=zipfile.ZIP_DEFLATED, compresslevel=6) as archive:\n"," if LORA_TEMPLATE_OUTPUT:\n"," for number, image_id in enumerate(complete_image_ids, 1):\n"," item = items_by_image_id[image_id]\n"," archive.write(os.path.join(EXPORT_DIR, f\"{image_id}.jpg\"), arcname=f\"{number}.jpg\")\n"," archive.writestr(f\"{number}_input.txt\", str(item.get(\"input_text\", \"\") or \"\"))\n"," archive.writestr(f\"{number}_output.txt\", str(item.get(\"output_text\", \"\") or \"\"))\n"," print(f\" {number}: triplet written\")\n"," else:\n"," for image_id in complete_image_ids:\n"," item = items_by_image_id[image_id]\n"," archive.write(os.path.join(EXPORT_DIR, f\"{image_id}.jpg\"), arcname=f\"{image_id}.jpg\")\n"," archive.writestr(f\"{image_id}.txt\", make_export_text(item))\n"," archive.write(JSONL_FILE, arcname=\"sankaku_lora_dataset.jsonl\")\n"," archive.writestr(\"search_metadata.txt\", search_metadata)\n","\n","print(\"ZIP:\", ZIP_FILE, f\"({os.path.getsize(ZIP_FILE)/1024**2:.2f} MB)\")\n","\n","results_df = pd.DataFrame(test_results)\n","for col in (\"image\", \"input_text\", \"output_text\", \"qwen_caption\", \"source\"):\n"," if col not in results_df.columns:\n"," results_df[col] = \"\"\n","results_df.to_csv(OUTPUT_FILE, index=False, encoding=\"utf-8\")\n","print(\"CSV:\", OUTPUT_FILE)\n","\n","# ============================================================\n","# HTML GALLERY\n","# ============================================================\n","\n","def download_sankaku_zip():\n"," if not os.path.isfile(ZIP_FILE):\n"," raise FileNotFoundError(\"ZIP not found.\")\n"," files.download(ZIP_FILE)\n","\n","output.register_callback(\"sankaku.download_zip\", download_sankaku_zip)\n","\n","def render_caption_tags(tags):\n"," rendered = []\n"," search_words = REQUIRED_TEST_TAGS + EXCLUDED_TEST_TAGS + ADDITIONAL_TEST_TAGS\n"," for tag in tags:\n"," tag_string = str(tag)\n"," safe = html.escape(tag_string)\n"," is_search = any(w in tag_string.lower() for w in search_words)\n"," cls = \"search-tag\" if is_search else \"normal-tag\"\n"," rendered.append(f'{safe}')\n"," return \" \".join(rendered)\n","\n","def make_copy_text(item):\n"," tags = item.get(\"tags\", [])\n"," if isinstance(tags, str):\n"," tags = tags.replace(\",\", \" \").split()\n"," tags_text = \" \".join(str(t).strip() for t in tags if str(t).strip())\n"," natural = normalize_caption_text(item.get(\"qwen_caption\", \"\"))\n"," if ENABLE_QWEN_CAPTIONER and natural:\n"," return (f\"{natural} {tags_text}\" if tags_text else natural), \"Copy Qwen caption + tags\"\n"," return tags_text, \"Copy tags\"\n","\n","def encode_copy_text(text):\n"," return base64.b64encode(text.encode(\"utf-8\")).decode(\"ascii\")\n","\n","cards = []\n","for item in test_results:\n"," image_id = str(item[\"id\"])\n"," image_uri = image_data_uris.get(image_id)\n"," if not image_uri:\n"," continue\n"," tags_html = render_caption_tags(item.get(\"tags\", []))\n"," rating = html.escape(str(item.get(\"rating\", \"\") or \"\"))\n"," meta_parts = []\n"," if pd.notna(item.get(\"width\")) and pd.notna(item.get(\"height\")):\n"," meta_parts.append(f\"{item.get('width')} × {item.get('height')}\")\n"," if pd.notna(item.get(\"fav_count\")):\n"," meta_parts.append(f\"Favourites: {item.get('fav_count')}\")\n"," if pd.notna(item.get(\"author_name\")) and str(item.get(\"author_name\") or \"\").strip():\n"," meta_parts.append(f\"Artist: {item.get('author_name')}\")\n"," source_label = item.get(\"source\", \"sankaku\")\n"," natural = normalize_caption_text(item.get(\"qwen_caption\", \"\"))\n"," natural_html = \"\"\n"," if natural:\n"," natural_html = f\"\"\"\n","
\n","
Qwen 3.5 caption
\n"," {html.escape(natural)}\n","
\"\"\"\n"," copy_text, button_label = make_copy_text(item)\n"," training_status = (\n"," \"LoRA JSONL training example\"\n"," if item.get(\"image\") and item.get(\"input_text\") and item.get(\"output_text\")\n"," else \"Not included in LoRA JSONL\"\n"," )\n"," cards.append(f\"\"\"\n","
![\"Sankaku](\"{image_uri}\")
\n","