swedish-print-ocr-3b (production)
Fine-tune of stanford-oval/churro-3B (itself a fine-tune of Qwen2.5-VL-3B-Instruct) for optical character recognition of Swedish printed books from 1600 to 1999, covering both blackletter (fraktur) and roman (antikva) typefaces.
This is the production model. It is trained on the full available mixture, including the 338 works that were held out to measure error rate elsewhere. It therefore has no valid measured CER of any kind — a number computed on memorized data is not a measurement. For a model with a genuine held-out CER, see Ericu950/swedish-print-ocr-3b-benchmark.
As reference, the benchmark model (same architecture and training recipe, held-out data excluded) reaches 3.25% CER (3.59 ± 0.31% over three seeds) on a 1,182-page, 214-work internal benchmark spanning 1600–1999, and 4.54% on an external Wikisource benchmark, after uniform decoding-repair applied identically to every autoregressive system compared. See the accompanying paper for full method, benchmark construction, and caveats — including known failure modes (repetition-loop decoding, page/column loss in parallel-text editions) that apply equally to this production checkpoint.
Data
Six sources at fixed proportions: NewsEye Swedish newspapers, Projekt Runeberg, GT4HistOCR (German), a Swedish fraktur corpus, Litteraturbanken, and Språkbanken regions — 34% roman Swedish, 44% Swedish blackletter, 22% German blackletter.
Training
Full fine-tune, frozen vision encoder, learning rate 8e-5, cosine decay, effective batch 128, bfloat16, ~1.76 epochs.
Usage
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
"Ericu950/swedish-print-ocr-3b", torch_dtype="bfloat16", device_map="auto"
)
processor = AutoProcessor.from_pretrained("Ericu950/swedish-print-ocr-3b")
License
Inherits the Qwen Research License Agreement from the base model chain (Qwen2.5-VL-3B → CHURRO-3B) — research use, see linked license. Not evaluated for commercial redistribution.
Limitations
Nothing about this checkpoint's own output has been proofread by a human at scale. See the paper's Limitations section for measurement pitfalls (loss ≠ CER, small validation sets, looping pages, page-alignment verification coverage) that apply to the training run this checkpoint came from.