Training
- Data: Riksarkivet's GT-Lejonet HTR ground truth (February 2026), the Swedish Lion's training side (its test split
test_0050 held 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 ┏━━━━━━━━┳━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓┃ subset ┃ lines ┃ mean CER ┃ CER (OpenBMB) ┃ WER (OpenBMB) ┃ mean CER, norm. ┃ cut off ┃ mean CER, loops cut ┃ CER (OpenBMB), loops cut ┃ loops cut ┃┡━━━━━━━━╇━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩│ all │ 6358 │ 6.67% │ 5.30% │ 21.12% │ 6.62% │ 0 │ 6.67% │ 5.30% │ 0 │└────────┴───────┴──────────┴───────────────┴───────────────┴─────────────────┴─────────┴─────────────────────┴──────────────────────────┴───────────┘ runs/htr/final/ood_minicpm.jsonl ┏━━━━━━━━┳━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓┃ subset ┃ lines ┃ mean CER ┃ CER (OpenBMB) ┃ WER (OpenBMB) ┃ mean CER, norm. ┃ cut off ┃ mean CER, loops cut ┃ CER (OpenBMB), loops cut ┃ loops cut ┃┡━━━━━━━━╇━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩│ all │ 6358 │ 7.21% │ 5.55% │ 21.76% │ 7.18% │ 0 │ 7.21% │ 5.55% │ 0 │└────────┴───────┴──────────┴───────────────┴───────────────┴─────────────────┴─────────┴─────────────────────┴──────────────────────────┴───────────┘ runs/htr/ckpt_eval/test_trocr.jsonl ┏━━━━━━━━┳━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓┃ subset ┃ lines ┃ mean CER ┃ CER (OpenBMB) ┃ WER (OpenBMB) ┃ mean CER, norm. ┃ cut off ┃ mean CER, loops cut ┃ CER (OpenBMB), loops cut ┃ loops cut ┃┡━━━━━━━━╇━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩│ all │ 5000 │ 4.17% │ 3.58% │ 14.61% │ 3.87% │ 0 │ 4.17% │ 3.58% │ 0 │└────────┴───────┴──────────┴───────────────┴───────────────┴─────────────────┴─────────┴─────────────────────┴──────────────────────────┴───────────┘ runs/htr/final/test_minicpm.jsonl ┏━━━━━━━━┳━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┓┃ subset ┃ lines ┃ mean CER ┃ CER (OpenBMB) ┃ WER (OpenBMB) ┃ mean CER, norm. ┃ cut off ┃ mean CER, loops cut ┃ CER (OpenBMB), loops cut ┃ loops cut ┃┡━━━━━━━━╇━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━┩│ all │ 5000 │ 5.22% │ 5.64% │ 20.06% │ 4.42% │ 3 │ 4.46% │ 3.92% │ 3 │└────────┴───────┴──────────┴───────────────┴───────────────┴─────────────────┴─────────┴─────────────────────┴──────────────────────────┴───────────┘
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 cut columns show what cutting such a tail
changes.
Licence
Apache-2.0, as its base model.