Fine-tuned TrOCR for French manuscript handwriting on Québec cadastral plans + notarial documents, trained as part of the GREF (Générateur de Reconstitution d'Extrait Foncier) pipeline.

Souverain, vérifiable, ré-entrainable. The training data, code, and decision report are public so any third party can re-run this exact model from scratch. See the Reproduction section below.

Training run

Base model: microsoft/trocr-large-handwritten @ main
Epochs: 3
Learning rate: 0.0002
Batch size: 16
Max target length: 64
Seed: 20260510
LoRA: on
Git SHA (worker repo): 05b32e5
Examples seen: 32,269
Wall-clock: 622.5s
Final training loss: 3.5342

Promotion gate

Verdict: ✅ promoted (tier: authentique)

Corpus

Mix recipe: default
Mix split: 0.95,0.025,0.025
Mix seed: 20260510

Role	sha256	rows	bytes
train	`57994aac20699beb…`	32,269	12,603,500
val	`fdb70c21e3c602ae…`	849	331,763
test	`238a349b9dfc3b35…`	849	331,789

Full per-file manifest (paths, full sha256, totals): see corpus_manifest.json.

Data sovereignty

Source crops + transcripts (the JSONL row contents) are never uploaded with this model. The corpus manifest records sha256 + row count of each JSONL file used by the run, which is enough for a third party to verify reproducibility against their own copy of the source corpora (CATMuS / HTR-United / BAnQ Wikisource / atelier human edits) but not enough to reconstruct the data. See docs/GREF_DATA_SOVEREIGNTY.md in the Fonciq monorepo for the full policy.

Reproduction

Run the canonical reproduction script that shipped with this run (also bundled at REPRODUCE.sh in this repo):

bash

#!/usr/bin/env bash
# Auto-generated by worker/scripts/gref_train_local.py at 2026-05-11T13:47:44Z.
#
# Reproduces the GREF HTR training run that produced
# ``pbel88/trocr-fr-cadastral-2026.05.bulletproof`` (or, when no push target was set, the run's final
# checkpoint). To use:
#
#   1. Clone the Fonciq monorepo (or your fork).
#   2. cd into the worker/ directory.
#   3. ./REPRODUCE.sh         (optional: WORK_DIR=/scratch/x ./REPRODUCE.sh)
#
# The script pins the repo at the exact git SHA that produced this run
# and installs the training-time deps from a fresh venv. It does not
# read your shell environment for credentials; set the relevant Hub /
# repository tokens in your shell yourself if you need push or private
# clone access.
#
# Source: docs/GREF_TRAIN_LAUNCH.md §"REPRODUCE.sh contract".

set -euo pipefail

REPO_ROOT="${REPO_ROOT:-$(git rev-parse --show-toplevel 2>/dev/null || pwd)}"
cd "$REPO_ROOT"
# 1. Pin the worker repo at the exact code that produced this run.
      if [ -n "${REPO_GIT_REMOTE:-}" ] && [ ! -e .git ]; then
        git clone --depth 50 "$REPO_GIT_REMOTE" .
      fi
      if git rev-parse --git-dir >/dev/null 2>&1; then
        git fetch --depth 50 origin "05b32e5" || true
        git checkout "05b32e5" || true
      fi

      # 2. Recreate the training-time Python environment from the pinned
      #    requirements (matches gref_train_launch.py's guaranteed baseline).
      python -m pip install --upgrade pip
      python -m pip install --extra-index-url https://download.pytorch.org/whl/cu124 \
        "torch>=2.6,<3"
      python -m pip install -r worker/requirements-train.txt

      # 3. Re-run the pipeline. WORK_DIR is parameterised so you can
      #    diff this run's outputs against the original.
      export PYTHONPATH=worker
      python worker/scripts/gref_train_local.py \
--work-dir ${WORK_DIR:-/tmp/gref-reproduce-$(date +%s)} \
--mix-recipe default \
--mix-split 0.95,0.025,0.025 \
--base-model microsoft/trocr-large-handwritten \
--revision main \
--epochs 3 \
--lr 0.0002 \
--batch-size 16 \
--max-target-len 64 \
--htr-united-src-id htru-lectaurep-bronod \
--banq-max-docs 5 \
--tier authentique \
--regression-tolerance 0.05 \
--catmus-max-rows 50000 \
--source htr-united \
--source wikisource-banq \
--source catmus \
--no-streaming \
--use-lora \
--push-to-hub pbel88/trocr-fr-cadastral-2026.05.bulletproof \
--git-sha 05b32e5

License

This model is released under the APACHE-2.0 license, matching the base TrOCR weights and every corpus in the training mix. Commercial use is permitted with attribution.