ExpertHTR Sparse MoE

Validation-selected page-level handwritten text recognition checkpoint from ExpertHTR. It is based on Qwen3.5-0.8B-Base and adds four routed experts plus one shared expert at layers 1, 5, 9, 13, 17 and 21.

Results

The selected checkpoint is training step 1914:

Metric	Value
Validation pages	962
Micro page CER	16.439%
Micro page WER	33.300%

CER removes only the leading region marker ([Rk]:). <del> and <gap> are kept as distinct OCR symbols.

Dataset

The companion gated dataset is available at 🤗 ExpertHTR-Dataset. The current dataset card excludes HWDB/CASIA. This checkpoint was trained on the original seven-source experiment, including HWDB2.0, so the companion dataset is not an exact reproduction of its training data. Do not describe this checkpoint as Bentham-only or claim that the gated dataset alone produced these weights. Follow every upstream dataset term; this model card does not grant rights to any training data.

Use with the repository

Clone the code, install dependencies, and follow the evaluation instructions:

git clone https://github.com/DAIR-Group/ExpertHTR
cd ExpertHTR
python -m pip install -e .
export HWVLM_DATA_ROOT=/path/to/dataset
export HWVLM_FINAL_TEST_CHECKPOINT=/path/to/this/model
experthtr final-test

The checkpoint uses ExpertHTR's custom full_model_state.pt format. Keep the tokenizer, processor, checkpoint_manifest.json, and the repository version together; do not load it as a generic Transformers checkpoint.

Files

This release contains runtime files only: model state, config, tokenizer, processor, chat template, checkpoint manifest, metadata and metrics. Training state, optimizer state, predictions, email addresses, access tokens and local machine paths are excluded. Review the manifest before mirroring this model.

License and intended use

The code is Apache-2.0. Qwen3.5-0.8B-Base and the training data retain their upstream licenses. This checkpoint is intended for research on handwritten text recognition and may make transcription errors. Do not use it for high-stakes decisions without human review.