Qwen3-VL-8B Hebrew — v19a (Cairo Genizah manuscript transcription)
Cairo Genizah AI Transcription — formal analysis and write-up of the project
The full account of the methods, training data, benchmarks and error analysis behind the Cairo Genizah AI transcription models. Search the transcribed corpus at Cairo Genizah AI.
Part of the Qwen3-VL-8B Hebrew / Cairo Genizah fine-tune series (https://github.com/AIStream-Peelout/historical-document-analysis). Flagship: v21b step 1200.
This version
Status: former flagship (August 2026), superseded by v20a and then v21b. Kept as the reference point
the later cards compare against.
- Trained 2026-08-20/21, 1,300 steps (the run stopped at ~1,310); warm start from v18b step 700
(weights only).
- What was new: the first KTIV data —
genizah_ktiv_v1 (914 manuscripts, 8,298 training rows of full
pages, prompted regions, sections and line crops) — and synthetic_hebrew_v3 (12k renders over twelve
faces: Solitreo, Rashi, STA"M and square-letter drills). Carried over: genizah_clean_v2 (1,055 curated
documentary pages with published transcriptions) and a small Talmud replay. Mixture: KTIV pages .30 /
documentary .20 / synthetic .20 / KTIV regions .08 / KTIV crops .07 / Talmud small-script .10 / gemara
crops .05.
- Recipe: the shared recipe with the merger frozen (
colab/genizah_focus_v19a.ipynb). Eval loss
0.781 → 0.554.
| benchmark |
step 1300 |
| religious CER* / F1 / readOK |
0.216 / 0.816 / 0.597 (two-column 0.255, single-column 0.205) |
| PGP CER* / F1 |
0.196 / 0.862 (Judaeo-Arabic F1 0.857) |
| grounding (zero-shot, no grounding data) |
locate 42/72 hits, median IoU 0.29; read-box CER 0.58 |
Canonical revisions: step 1300 43e21bd7a6fedd323879fc5ea2c298df90a92784; step 700
10c8cbf85e7ccaaba95febda37fa0d173c83b4ea.
The series at a glance
All versions are QLoRA adapters on Qwen/Qwen3-VL-8B-Instruct (r16/α16 on the language model and every
vision-tower block, merger frozen unless stated; batch 1 × 8, AdamW-8bit, LR 5e-5 cosine, image budget
6.5–7 MP, eval/save every 100 steps), trained on Colab A100s and evaluated with one decode at T=0.1
through an 8-bit MLX conversion. Two benchmarks recur below: religious = 140 held-out KTIV Talmud
pages (33 two-column), PGP = 131 held-out documentary fragments from the Princeton Geniza Project
(82 Judaeo-Arabic). CER* is the median character error rate over substantive outputs; page-level flips
between versions are within decode noise, medians are the claim. The flagship is v21b step 1200.
| version |
trained |
data change |
recipe change |
religious CER* |
PGP CER* |
status |
| v19a |
2026-08-20/21 |
+ KTIV v1 (8.3k rows), synthetic Hebrew v3 |
base recipe, merger frozen |
0.216 |
0.196 |
former flagship, superseded |
| v19b |
2026-08-28/29 |
= v19a |
merger r16 LoRA |
0.234 |
0.186 |
control, not adopted |
| v19c |
2026-08-30 → 09-01 |
= v19a |
merger trained full-weight |
0.225 |
0.189 |
not adopted |
| v20a |
2026-09-03/05 |
KTIV v2 (2.7× letters) + 8 % grounding |
warm start v19a-1300 |
0.184 |
0.183 |
superseded; merged weights published |
| v20b |
2026-09-06/09 |
= v20a |
merger r16 LoRA |
n/a |
n/a |
dead-end arm (boxes worse) |
| v21 |
2026-09-11 |
KTIV v3, 20 % grounding, streamed |
warm start v20a-1800 |
— |
— |
scrap: trained without images |
| v21b |
2026-09-11/15 |
= v21 |
dispatch fix + dataloader gate |
0.165 |
0.177 |
flagship / model of record |
| v22a |
2026-09-25/26 |
+ PGP editions, documentary grounding, extractive QA; no synthetic / Talmud |
warm start v21b-1200, map-style data |
0.186 |
0.164 |
pilot; documentary best, religious regression |
Baselines on the same tables: the best Kraken setup reads religious pages at CER 0.30 and PGP at 0.25;
stock Qwen3-VL-8B produces a substantive transcription on 2 % of PGP pages (F1 0.01).
Training data comes from the KTIV project of the National Library of Israel (literary manuscripts and
transcriptions), from published editions of Princeton Geniza Project documents, and from synthetic renders;
see each card's data section. Code, builders and the evaluation harness: https://github.com/AIStream-Peelout/historical-document-analysis.
Loading
The repo root holds the adapter of the latest saved step; every step is a commit ("step N" = root
adapter, "step N, checkpoint" = the same plus last-checkpoint/ with optimizer state). Pin the revision
you want:
from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel
base = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3-VL-8B-Instruct", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "isaacmg/qwen3-vl-8b-hebrew-v19a-ckpt", revision="43e21bd7a6fedd323879fc5ea2c298df90a92784") # step 1300
processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-8B-Instruct", min_pixels=6_500_000, max_pixels=7_000_000)
Prompt the model the way it was trained: one page image plus a plain instruction such as
"Transcribe this manuscript page line by line." For the grounding tasks (from v20a on) use the exact
prompts in the builders (src/finetuning/qwen_hebrew/build_ktiv_dataset.py); boxes are [x1, y1, x2, y2]
normalised to 0–1000 of the original image.
Data rights
Manuscript images remain the property of their holding institutions; the KTIV transcriptions are the
National Library of Israel's; the Princeton Geniza Project's editions are used under their research terms;
the Talmud replay set (v19a–v21b) consists of scans marked for non-commercial use. This adapter is released
for research on handwritten-text recognition. Credit NLI KTIV and the Princeton Geniza Project when
you build on it.
