Qwen3-VL-8B Hebrew — v19c (merger trained full-weight; not adopted)
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: not adopted. The second arm of the merger ablation; use v21b.
- Trained 2026-08-30 → 09-01 in two Colab sessions (steps 100–1200, then 1300–1800; never reached
2,000); warm start v18b step 700; data identical to v19a.
- The one change: the merger (~160M parameters) trains at full weight via
modules_to_save
(colab/genizah_focus_v19c.ipynb). Best eval loss of the v19 generation: 0.544 at 1200, 0.531 at 1800.
| benchmark |
step 1800 |
step 1200 |
| religious CER* / F1 / readOK |
0.225 / 0.811 / 0.590 (two-column 0.355) |
0.218 (two-column 0.537) |
| PGP CER* / F1 |
0.189 / 0.870 |
0.190 / 0.867 |
| grounding (zero-shot) |
locate 39/72, IoU 0.32; read-box CER 0.77; grounded line IoU 0.36 |
Finding: the full-weight merger cured the pages v19b broke but each checkpoint broke a different
handful of pages instead, and medians only tied v19a. Together with v19b this fixed the series decision:
capacity in the merger relocates instability rather than removing it, so the merger is frozen from v20 on.
Canonical revisions: step 1800 6620fb0da7639c66f20195bef4463bce59ba810e; step 1200
34e7d8d804e624a9c08fff961975579f1fb0c739.
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-v19c-ckpt", revision="6620fb0da7639c66f20195bef4463bce59ba810e") # step 1800
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.
