lector-4b-v1

A page reader for Latin scholastic print (16th–20th century): one page image in, one Markdown page file out, in the diplomatic-v1 dialect (print spelling, æ/œ, &, accents, u/v, abbreviation marks; long s as s; footnotes ^[…], marginal notes ^[⟦M⟧ …], a footnote continued from the previous page as a last block starting ⟦CONTINUED-NOTE⟧). Full fine-tune of Qwen/Qwen3.5-4B (4.66B parameters, bf16 9.3 GB), 2 epochs, 3,732 pages + 2,000 lines, by lector (Schola Thomistica), 2026-11-01; internal run ft6. Same data and prompt as lector-2b-v1: only the size differs.

The name: a lector read the set text aloud and expounded it in the medieval schools; the Dominican order still grants the degree of Lector of Sacred Theology.

Use it exactly like this

• Prompt: prompt.txt in this repo, verbatim, as the only text of one user turn after the image. It is the prompt the model was trained on; later lector prompts belong to later models.
• Image: one portrait page (split two-page spreads at the gutter first), RGB, longest side resized to 1280 px (Lanczos) when larger. Processor: min_pixels = 256*32*32, max_pixels = 1280*1280.
• Decoding: bf16, greedy (do_sample=False, repetition_penalty=1.0), max_new_tokens=3500, stop at <|im_end|> or eos, chat template with enable_thinking=False.
• Loop guard on the output (lector prompts.unloop), always:

  text = re.sub(r"^\|.*$\n?", "", text, flags=re.M)                 # tables are not in the dialect
  text = re.sub(r"(.{12,1500}?)\1{2,}", r"\1", text, flags=re.S)    # a run repeated 3+ times -> once
  

  A page where the guard fired lost text after the loop: re-read it another way.

Scores (lector-bench v0, md_cer folded, page-weighted, through the loop guard)

	all 146 pages	public split, 86 pages	public: omit30 / insert30	marginal notes found
ft2 (Qwen3-VL-2B, internal)	0.137	0.059	8 / 1	4 / 103
ft4 (Qwen3.5-2B, internal)	0.099	0.089	14 / 3	45 / 103
lector-2b-v1 (ft5)	0.092	0.066	6 / 3	48 / 103
lector-4b-v1 (ft6)	0.076	0.039	3 / 1	71 / 103

omit30 / insert30: pages where a run of 30+ words was skipped or invented. Álvarez 1610: 0.022 (2b-v1 0.053); Báñez 1602: 0.256 (0.449); 19th–20th-century print: 0.002–0.04. Validation loss 0.087 (2b-v1 0.103). Generation took about as long per page as 2b-v1 (≈ 8 s; 2b-v1 on an H100, this one on an H200).

Known weaknesses

• Very dense, heavily abbreviated old print (Báñez 1602: 0.26), which no training page resembles.
• It can stop at the foot of the first column of a two-column old page (Hurtado p. 42: half the page). Check the length of what it returns against the page.
• Loops are rarer than in 2b-v1 but possible; keep the guard.
• It sometimes writes ^[⟦M⟧⟦R⟧ …] (paragraph summary). Treat that role as a guess (2b-v1's precision was 0.59 on v0): the kind of a marginal note is decided at scouting/QA, not by the reader.