Isaac Godfried

Part of the Qwen3-VL-8B Hebrew / Cairo Genizah fine-tune series (https://github.com/AIStream-Peelout/historical-document-analysis). Flagship: v21b step 1200. Trained 2026-09-06/09 (steps 100–200, then 300–2000); the only run in the series that completed its 2,000 steps. Warm start v19a step 1300; data identical to v20a, so it is a clean A/B. merger is the lever for better boxes. Eval loss 0.750 → 0.668. Finding: merger LoRA made the box geometry slightly worse and the transcription no better; the lever for accurate boxes is direct grounding data, not merger capacity. Final step 2000 4cae9d52d9cd167cb0d1ddd0248ed728c7308f29. All versions are QLoRA adapters on Qwen/Qwen3-VL-8B-Instruct…

Part of the Qwen3-VL-8B Hebrew / Cairo Genizah fine-tune series (https://github.com/AIStream-Peelout/historical-document-analysis). Flagship: v21b step 1200. 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. families), curated documentary pages.165, synthetic.13, Talmud replay.12. (colab/genizahv21bdispatchfix.ipynb). Eval loss 0.697 → 0.655 at step 1200 (best); 0.657 at 1700. Why it is the flagship: the two-column penalty that…

Part of the Qwen3-VL-8B Hebrew / Cairo Genizah fine-tune series (https://github.com/AIStream-Peelout/historical-document-analysis). Flagship: v21b step 1200. Status: pilot, not promoted. Best documentary reader in the series; a step back on religious pages; QA not yet learned. The full v22 run will follow with several times more unique rows. Trained 2026-09-25/26, steps 100–1300 (the Colab session hit its 24-hour cap at ~1,340). Warm start v21b step 1200. training rows, 188 validation rows on pages disjoint from training, 4,598 images): KTIV v4 transcription.30, line-broken PGP editions.25, KTIV v4 grounding.15, documentary grounding.10 (Kraken ink boxes on lines where Kraken and v21b…

Part of the Qwen3-VL-8B Hebrew / Cairo Genizah fine-tune series (https://github.com/AIStream-Peelout/historical-document-analysis). Flagship: v21b step 1200. 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): genizahktivv3 (60,001 rows, 3,700 pages / 1,588 manuscripts, 2.2M letters) and six direct-grounding families — locateword, readboxword, lineindex, lineofphrase, groundeddetect (box before text) and groundedcrop — lifting grounding to 20 % of the mixture across ten families. The dataset was streamed per task split. sliced the batch tensors…

Part of the Qwen3-VL-8B Hebrew / Cairo Genizah fine-tune series (https://github.com/AIStream-Peelout/historical-document-analysis). Flagship: v21b step 1200. Status: superseded by v21b (which warm-starts from this checkpoint). Merged bf16 weights of step 1800 are published as isaacmg/qwen3-vl-8b-hebrew-v20a-merged. Trained 2026-09-03/05 in two sessions (steps 100–600, then 700–1800 until the session cap); warm start v19a step 1300. transcribed letters (2.7× v19a's KTIV slice) with repaired layout ground truth — and the first grounding and layout families with boxes in 0–1000 coordinates: locate (6,613 rows), readbox (3,242), layoutqa (3,301), groundedpage (268), at an 8 % share of the…

Part of the Qwen3-VL-8B Hebrew / Cairo Genizah fine-tune series (https://github.com/AIStream-Peelout/historical-document-analysis). Flagship: v21b step 1200. 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). pages, prompted regions, sections and line crops) — and synthetichebrewv3 (12k renders over twelve faces: Solitreo, Rashi, STA"M and square-letter drills). Carried over: genizahcleanv2 (1,055 curated documentary pages with published transcriptions) and a small Talmud replay. Mixture: KTIV…

Merged bf16 weights of the v2.0a checkpoint (v20a-ckpt, step 1800, commit af9df6a0): the LoRA adapters (language layers r16, vision tower) folded into Qwen3-VL-8B-Instruct. Ready for transformers, vLLM, or conversion. The image processor ships the model's resolution contract (minpixels=6500000, maxpixels=7000000): send the largest image you have and let the processor resize it. Status: current best of the series on both transcription benchmarks (Sep 2026) and the first checkpoint with text grounding. Serving demo: the Genizah Reader Space. Diplomatic transcription of handwritten Hebrew-script manuscripts (Hebrew, Aramaic, Judaeo-Arabic) — Cairo Genizah documentary fragments and…

Unless you are studying this series' training history, use the flagship instead: qwen3-vl-8b-hebrew-v19a-ckpt. Status: superseded (historical). The Rashi-script generation: warm-started from the v1 winner and trained on a mixture led by synthetic Rashi-script data (40%) plus Talmud small-script and gemara material, at a 3.2 MP resolution floor. on a held-out synthetic Rashi eval slice whose text is unmemorizable, i.e. the model reads glyphs rather than reciting canonical text. Also solved crop-level Vilna gemara (CER 0.019). - Talmud page-level (strict CER means, 65-page benchmark): gemara 0.059 / rashi 0.362 / tosafot 0.512. One line per generation — which checkpoint to use and which are…

Part of the Qwen3-VL-8B Hebrew / Cairo Genizah fine-tune series (https://github.com/AIStream-Peelout/historical-document-analysis). Flagship: v21b step 1200. 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. (colab/genizahfocusv19c.ipynb). Best eval loss of the v19 generation: 0.544 at 1200, 0.531 at 1800. 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…

This is a historical merged export, NOT the current flagship. The flagship is v1.9a — adapter + merge instructions at qwen3-vl-8b-hebrew-v19a-ckpt. Full 16-bit merge of v1.8b step 700 — the first generation whose vision tower actually trained (patch-embed gradient fix), and the warm-start ancestor of all v1.9 models. 6.5 MP resolution policy ships in the processor config. Superseded by v1.9a for all use. One line per generation — which checkpoint to use and which are historical: \\ CER over substantive attempts only, alignment-based scorer. Benchmarks differ across generations (Talmud print vs Genizah manuscripts) — compare within a row's named benchmark, not across rows.

This is a historical merged export, NOT the current flagship. The flagship is v1.9a — adapter + merge instructions at qwen3-vl-8b-hebrew-v19a-ckpt. Full 16-bit merge of v1.5, the Rashi-script generation: pure Rashi glyph perception proven (CER 0.018 on unmemorizable synthetic text), crop-level Vilna gemara solved (CER 0.019), Talmud page-level gemara CER 0.059. Trained at a 3.2 MP resolution floor — use that policy at inference. Vision tower unadapted. Superseded by v1.6+ on print and v1.9a on manuscripts. One line per generation — which checkpoint to use and which are historical: \\ CER over substantive attempts only, alignment-based scorer. Benchmarks differ across generations (Talmud…

Unless you are studying this series' training history, use the flagship instead: qwen3-vl-8b-hebrew-v19a-ckpt. Status: superseded — A/B control arm. v1.8 was a controlled experiment on whether training the vision tower matters for handwriting: v1.8a kept the vision LoRA frozen (the accidental status quo of v1–v1.7, here made deliberate), while its twin v1.8b trained it. v1.8b won the comparison and became the ancestor of every later checkpoint. Keep this repo for reproducing the ablation; do not use it for inference. One line per generation — which checkpoint to use and which are historical: \ CER over substantive attempts only, alignment-based scorer. Benchmarks differ across generations…

This is a historical merged export, NOT the current flagship. The flagship is v1.9a — adapter + merge instructions at qwen3-vl-8b-hebrew-v19a-ckpt. Full 16-bit merge of the series' first-generation (v1) Talmud adapter — the first model in this line to read Vilna-print gemara from pixels. Vision tower unadapted (language-only LoRA generation). Superseded on every (rope under textconfig.ropeparameters); MLX conversion requires patching ropetheta to top level. One line per generation — which checkpoint to use and which are historical: \\ CER over substantive attempts only, alignment-based scorer. Benchmarks differ across generations (Talmud print vs Genizah manuscripts) — compare within a…

Unless you are studying this series' training history, use the flagship instead: qwen3-vl-8b-hebrew-v19a-ckpt. Status: superseded. The first generation trained on real Cairo Genizah handwriting (documentary fragments) alongside the Talmud-print and synthetic material — the series' pivot from print to manuscripts. open VLM we measured on both our Talmud-print and Genizah-manuscript benchmarks. v1.8b). Later generations (v1.8b → v1.9a) supersede it on every manuscript benchmark. One line per generation — which checkpoint to use and which are historical: \ CER over substantive attempts only, alignment-based scorer. Benchmarks differ across generations (Talmud print vs Genizah manuscripts)…

Unless you are studying this series' training history, use the flagship instead: qwen3-vl-8b-hebrew-v19a-ckpt. reference. The full-page generation: page-level training at the 6.5 MP resolution policy the series has used ever since. Talmud page-level strict CER means (65-page benchmark): gemara 0.090 / rashi 0.047 / tosafot 0.099 — the Rashi-script column 6.7× better than the strongest closed model we evaluated (0.315). - Step 1100, b9f47f32e9ed51eb45598e27ffb5c5b650f7723c (run end): the v1.7 warm start. - Medians across checkpoints were ceiling-level (≈0.004–0.013 per section); checkpoint ranking is driven by collapse-page counts — see the flagship card's notes on failure-mode-aware…

This is a historical merged export, NOT the current flagship. The flagship is v1.9a — adapter + merge instructions at qwen3-vl-8b-hebrew-v19a-ckpt. Full 16-bit merge of v1.8a, the vision-frozen control arm of the v1.8 A/B experiment (its twin v1.8b trained the vision tower and won). Kept for reproducing the ablation; do not use for inference. One line per generation — which checkpoint to use and which are historical: \\ CER over substantive attempts only, alignment-based scorer. Benchmarks differ across generations (Talmud print vs Genizah manuscripts) — compare within a row's named benchmark, not across rows.

This model is a fine-tuned version of unsloth/qwen3-vl-8b-instruct-unsloth-bnb-4bit. It has been trained using TRL. This model was trained with SFT.

A kraken 7 baseline-segmentation model fine-tuned on Cairo Genizah manuscript pages — the first Genizah-specific segmenter we know of that is publicly available. Recognition models for these hands exist (the MiDRASH project published its recognizers on Zenodo), but line segmentation is where a generic model loses most of the text on Genizah fragments: tight columns, damaged edges, script far larger than the segmenter was trained on, marginal glosses and rotated scans. This model closes a large part of that gap. 10.5281/zenodo.22879549, Apache-2.0), a general print-and-handwriting segmenter trained on Latin, Arabic and Greek script corpora. Cairo Genizah, drawn from the KTIV project of the…