Qwen3-VL-8B Hebrew — v21b (flagship: KTIV v3 + direct grounding)
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: flagship and model of record (step 1200). It is the reader behind the two-reader consensus
pipeline on cairogenizah.ai and the warm start of v22a.
- Trained 2026-09-11/15: steps 100–1200 in one session (the session died at ~1,240), then a
continuation 1300–1700 (died at ~1,780). Warm start v20a step 1800.
- Data: as v21 —
genizah_ktiv_v3 plus the six direct-grounding families (grounding 20 %, ten
families), curated documentary pages .165, synthetic .13, Talmud replay .12.
- Recipe: v21's with the dispatcher fix, the prepared-dataloader gate and stream-skipping resume
(
colab/genizah_v21b_dispatchfix.ipynb). Eval loss 0.697 → 0.655 at step 1200 (best); 0.657 at 1700.
| benchmark |
step 1200 |
step 1700 |
| religious CER* / F1 / readOK |
0.165 / 0.869 / 0.692; two-column 0.164 = single-column 0.166 |
0.175 / 0.864 (two-column 0.208) |
| PGP CER* / F1 |
0.177 / 0.880 (Judaeo-Arabic 0.872, Hebrew 0.897) |
0.177 / 0.883 |
| grounding |
locate 56/72 hits, median IoU 0.49; read-box CER 0.16; grounded line IoU 0.62, boxes on the right line 80 % |
Why it is the flagship: the two-column penalty that every earlier version paid is gone, boxes land on
the right line four times in five, and it is the best religious reader in the series. Known limits: on
documentary photographs the row boxes stretch beyond the ink (about 43 % of rows sit below it); pages with
two text blocks are often read as one; word-level tasks only match v20a.
Canonical revisions: step 1200 6724c32cd0f6296223978ae69635fc782d2dbb03; step 1700
e9b3b28fa7752cfd20421d0a395cd674cb889b17.
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-v21b-ckpt", revision="6724c32cd0f6296223978ae69635fc782d2dbb03") # step 1200
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.
