PaddleOCR-VL-1.6 is a plug-and-play 0.9B upgrade of 1.5 (96.33% SOTA on
OmniDocBench v1.6). vLLM offline backend with the six task modes (ocr, table,
formula, chart, spotting, seal) and --config/--create-pr for benchmark runs.
Must run in image-mode on HF Jobs (vllm/vllm-openai image) so flashinfer's
CUDA kernels are reused; the default uv-script image has no nvcc. Verified
end-to-end on a100-large: 5/5 clean markdown on ufo-ColPali, ~240 tok/s.
nuextract3: add --instructions for chat_template_kwargs.instructions
Exposes NuExtract3's `instructions` chat-template kwarg, a free-text
steering lever for output formats and routing across union/optional
schema branches. Mirrors the existing `--enable-thinking` passthrough:
new arg, plumbed via main() into chat_template_kwargs only when truthy
(no-op when omitted).
Docs: https://huggingface.co/numind/NuExtract3#instructions
Smoke (davanstrien/ufo-ColPali, 3 samples, extraction mode,
template={"title": "string", "author": "string"},
instruction="Prepend 'INSTR-OK::' to the title value"):
sentinel landed in 3/3 extractions.
Teklia French vital-records (58 cards, superset Baptême/Mariage/Décès
template + type_acte enum + French routing instruction): strict
F1 0.281 -> 0.321. Modest; the model partially honours the
branch-null routing (emits empty {} subdicts rather than null for
non-matched branches), which is consistent with NuExtract3's
shape-binding template engine. Per-user instruction tuning lives
downstream of this flag.
dots-mocr: skip whole batch when an image fails to decode
davanstrien+1: Claude Opus 4.7 (1M context)HF Staff
1bb77ce
Same class of bug as pp-doclayout's (15cd760) and smoldocling's (8b99222)
fixes from yesterday: the list comprehension
`[dataset[i][image_column] for i in batch_indices]` lives outside the
existing try/except, so one PIL.UnidentifiedImageError takes down the
entire run.
Hit by an overnight 5k BHL run on l4x1 — crashed at batch 150/313 (48%)
after ~10h on a corrupt JP2, no output saved.
Now wrapped in its own try/except. On a corrupt image, the whole batch
is marked `[OCR SKIPPED — UNREADABLE IMAGE]` (length-preserving so the
final add_column still aligns) and the run continues.
Smoke-tested via hf jobs uv run on a 50-image clean slice — confirmed
happy path still works end-to-end (14.2 min, dataset pushed cleanly).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
