# Reproducing the conversion

These instructions reproduce the publication inputs from pinned upstream revisions. They intentionally download the
original checkpoint from Hugging Face and do not bypass gated access or licensing.

## Inputs

- `datalab-to/chandra-ocr-2` at `af93b47dba1b47b6640c86ccf487ed2260ab9a09`
- `ggml-org/llama.cpp` at `8f5ab832ca7d8a7b4f23687693fb8b0ecbc227e7`
- Python 3.12

The pinned source `model.safetensors` must hash to:

```text
0804568be9f099d6479fad9ed77a4da4611f3c1e7bc6e009af7dce45e8aa3847
```

## Commands

```bash
git clone https://github.com/ggml-org/llama.cpp.git
git -C llama.cpp checkout --detach 8f5ab832ca7d8a7b4f23687693fb8b0ecbc227e7

python3.12 -m venv llama.cpp/.venv-convert
llama.cpp/.venv-convert/bin/pip install \
  -r llama.cpp/requirements/requirements-convert_hf_to_gguf.txt \
  'huggingface_hub[cli]'

llama.cpp/.venv-convert/bin/hf download datalab-to/chandra-ocr-2 \
  --revision af93b47dba1b47b6640c86ccf487ed2260ab9a09 \
  --local-dir chandra-ocr-2-source

sha256sum chandra-ocr-2-source/model.safetensors

llama.cpp/.venv-convert/bin/python llama.cpp/convert_hf_to_gguf.py \
  chandra-ocr-2-source \
  --outfile chandra-ocr-2.BF16.gguf \
  --outtype bf16 \
  --no-mtp

llama.cpp/.venv-convert/bin/python llama.cpp/convert_hf_to_gguf.py \
  chandra-ocr-2-source \
  --outfile chandra-ocr-2.mmproj-bf16.gguf \
  --outtype bf16 \
  --mmproj

sha256sum chandra-ocr-2.BF16.gguf chandra-ocr-2.mmproj-bf16.gguf
```

The converter can add its `mmproj-` prefix depending on the exact output name. Rename the generated projector to
`chandra-ocr-2.mmproj-bf16.gguf` before comparing it with the published manifest.

## Expected outputs

```text
4e9d5fa9854cf820d4425d28034df31ec1221a7f9d1082b0c4359d79f318cb56  chandra-ocr-2.BF16.gguf
54ddb8285933512cdbf1c84238aa0435b473a6efef2caeda8ca802c2899e87b3  chandra-ocr-2.mmproj-bf16.gguf
```

If the hashes differ, retain the new converter revision, complete command line, source revision, and resulting hashes
instead of relabeling the output as this build.