diffu-minicpm-ocr-htr
A full fine-tune of MiniCPM-OCR-1.0 that reads Swedish handwritten archive records: one text line, or a whole text region (its lines in reading order, one to a line). It is the reader of diffu, trained to be compared with Riksarkivet's Swedish Lion TrOCR on the same data and the same test lines.
Its proof-reader is diffu-minicpm-ocr-htr-jev and its DSpark draft for faster decoding diffu-minicpm-ocr-htr-dspark, where they have been published.
Training
- Data: Riksarkivet's GT-Lejonet HTR ground truth (February 2026), the Swedish Lion's training side (its test split
test_0050held out): 1,481,877 text lines, each crop white outside its line's polygon, and 83,072 text regions, each crop masked the same way. Lines whose text marks an uncertain reading (??,(?),[,]) and lines under 10 pixels are left out, as the baseline left them out. The text is trained as transcribed, not normalised. - Recipe: OpenBMB's full fine-tune (vision encoder and language model trained, the aligner frozen), learning rate 5e-6 on a cosine schedule planned over 3 epochs, 128 samples a step, bfloat16, on two RTX PRO 6000 Blackwell GPUs. Stopped after 28,000 of its 36,681 steps (2.3 epochs), where the held-out lines' error had all but stopped falling.
- Checkpoint: checkpoint-28000, the one of lowest corpus CER on 1,000 held-out dev lines.
Use
Needs vLLM at commit fa3b9d13 or later (MiniCPM-OCR support, vllm#64105),
as OpenBMB's installer builds it (OpenBMB/MiniCPM-OCR). Crop as in training:
a line's or region's bounding box, white outside its polygon, scaled down to at most 448 x 448 pixels' worth (200,704;
a region up to 9 times that).
vllm serve Gabriel/diffu-minicpm-ocr-htr --served-model-name htr --max-model-len 8192 \
--limit-mm-per-prompt '{"image": 4, "video": 0}' \
--mm-processor-kwargs '{"downsample_mode": "4x", "max_slice_nums": 9}' --trust-request-mm-kwargs \
--default-chat-template-kwargs '{"enable_thinking": false}'
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="-")
reply = client.chat.completions.create(
model="htr",
messages=[{"role": "user", "content": [
{"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,..."}},
{"type": "text", "text": "Text Recognition:"},
]}],
max_tokens=512, # a line; a region takes up to 4096
temperature=0.01, top_p=0.001, extra_body={"top_k": 1, "repetition_penalty": 1.0},
)
print(reply.choices[0].message.content)
Only Text Recognition: was trained. A reading that ends with finish_reason == "length" has looped (dotted leader
lines do); cut its repeated tail, as the loops cut columns below do.
Evaluation
This checkpoint read through vLLM as above, beside the Swedish Lion TrOCR's greedy readings of the same lines: 5,000
lines of the baseline's own test split (test_0050, test_* rows) and all 6,358 lines of
Riksarkivet/eval_htr_out_of_domain_lines
(ood_* rows). Each table is one reader on one set (trocr the baseline, minicpm this model).
- mean CER: the per-line mean, whitespace normalised (the baseline's own scoring).
- CER, WER (OpenBMB): corpus rates, all edits over all reference characters or words (NFC, ends trimmed).
- mean CER, norm.: the baseline's training normalisation applied to both sides.
- cut off: readings that reached the token limit.
- loops cut: each reading scored again with a repeating tail cut to one copy.
runs/htr/eval/ood_trocr_greedy.jsonl
ββββββββββ³ββββββββ³βββββββββββ³ββββββββββββββββ³ββββββββββββββββ³ββββββββββββββββββ³ββββββββββ³ββββββββββββββββββββββ³βββββββββββββββββββββββββββ³ββββββββββββ
β subset β lines β mean CER β CER (OpenBMB) β WER (OpenBMB) β mean CER, norm. β cut off β mean CER, loops cut β CER (OpenBMB), loops cut β loops cut β
β‘βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ©
β all β 6358 β 6.67% β 5.30% β 21.12% β 6.62% β 0 β 6.67% β 5.30% β 0 β
ββββββββββ΄ββββββββ΄βββββββββββ΄ββββββββββββββββ΄ββββββββββββββββ΄ββββββββββββββββββ΄ββββββββββ΄ββββββββββββββββββββββ΄βββββββββββββββββββββββββββ΄ββββββββββββ
runs/htr/final/ood_minicpm.jsonl
ββββββββββ³ββββββββ³βββββββββββ³ββββββββββββββββ³ββββββββββββββββ³ββββββββββββββββββ³ββββββββββ³ββββββββββββββββββββββ³βββββββββββββββββββββββββββ³ββββββββββββ
β subset β lines β mean CER β CER (OpenBMB) β WER (OpenBMB) β mean CER, norm. β cut off β mean CER, loops cut β CER (OpenBMB), loops cut β loops cut β
β‘βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ©
β all β 6358 β 7.21% β 5.55% β 21.76% β 7.18% β 0 β 7.21% β 5.55% β 0 β
ββββββββββ΄ββββββββ΄βββββββββββ΄ββββββββββββββββ΄ββββββββββββββββ΄ββββββββββββββββββ΄ββββββββββ΄ββββββββββββββββββββββ΄βββββββββββββββββββββββββββ΄ββββββββββββ
runs/htr/ckpt_eval/test_trocr.jsonl
ββββββββββ³ββββββββ³βββββββββββ³ββββββββββββββββ³ββββββββββββββββ³ββββββββββββββββββ³ββββββββββ³ββββββββββββββββββββββ³βββββββββββββββββββββββββββ³ββββββββββββ
β subset β lines β mean CER β CER (OpenBMB) β WER (OpenBMB) β mean CER, norm. β cut off β mean CER, loops cut β CER (OpenBMB), loops cut β loops cut β
β‘βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ©
β all β 5000 β 4.17% β 3.58% β 14.61% β 3.87% β 0 β 4.17% β 3.58% β 0 β
ββββββββββ΄ββββββββ΄βββββββββββ΄ββββββββββββββββ΄ββββββββββββββββ΄ββββββββββββββββββ΄ββββββββββ΄ββββββββββββββββββββββ΄βββββββββββββββββββββββββββ΄ββββββββββββ
runs/htr/final/test_minicpm.jsonl
ββββββββββ³ββββββββ³βββββββββββ³ββββββββββββββββ³ββββββββββββββββ³ββββββββββββββββββ³ββββββββββ³ββββββββββββββββββββββ³βββββββββββββββββββββββββββ³ββββββββββββ
β subset β lines β mean CER β CER (OpenBMB) β WER (OpenBMB) β mean CER, norm. β cut off β mean CER, loops cut β CER (OpenBMB), loops cut β loops cut β
β‘βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ©
β all β 5000 β 5.22% β 5.64% β 20.06% β 4.42% β 3 β 4.46% β 3.92% β 3 β
ββββββββββ΄ββββββββ΄βββββββββββ΄ββββββββββββββββ΄ββββββββββββββββ΄ββββββββββββββββββ΄ββββββββββ΄ββββββββββββββββββββββ΄βββββββββββββββββββββββββββ΄ββββββββββββ
This model read the 5,000 test lines at 68.55 lines a second on one RTX PRO 6000 Blackwell (vLLM, 64 requests at a time); the baseline read them at 7.91 (transformers, float32, batches of 32, greedy).
Limitations
- Trained and tested on GT-Lejonet's collections. Hands, periods and kinds of record it has not seen read worse (the out-of-domain rows).
- Like other vision-language readers, it can write a familiar modern spelling for an old one, or drop a letter, where a dedicated line recogniser copies the strokes.
- Near-greedy decoding can repeat one unit to the token limit; the
loops cutcolumns show what cutting such a tail changes.
Licence
Apache-2.0, as its base model.