Qwen3-VL-8B Hebrew — v21 (image-blind run; do not use)
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: scrap, kept as evidence. Never resume from it; use v21b, the corrected re-run.
- Trained 2026-09-11, steps 100–700, then stopped. Warm start v20a step 1800.
- What was new (and carried into v21b):
genizah_ktiv_v3 (60,001 rows, 3,700 pages / 1,588
manuscripts, 2.2M letters) and six direct-grounding families — locate_word, read_box_word, line_index,
line_of_phrase, grounded_detect (box before text) and grounded_crop — lifting grounding to 20 % of the
mixture across ten families. The dataset was streamed per task split.
- What went wrong: with a streamed (iterable) dataset, Accelerate's default dataloader dispatcher
sliced the batch tensors on their first dimension, cutting
pixel_values from [25728, 1536] to
[1, 1536]. The model trained on text alone: train loss 3.20 at step 10 (v20a: 0.51) and eval loss rising
from 0.683 to 0.764 against a 0.673 warm start. Full write-up: docs/postmortem_v21_dispatch_truncation.md.
- Lesson baked into every later notebook:
accelerator_config={"dispatch_batches": False} and a gate
that inspects a batch from the prepared dataloader (shapes against the image grid, prepared loss equal
to direct loss) before step 1.
No benchmark results; step 700 636da9c1ad881a17389ca9ec901963f5e06101cf exists only for the postmortem.
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-v21-ckpt", revision="636da9c1ad881a17389ca9ec901963f5e06101cf") # step 700
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.
