README.md

metadata

title: Ghostwriter API
emoji: 👻
colorFrom: purple
colorTo: gray
sdk: docker
app_file: api.py
pinned: false

Ghostwriter 👻📓

Fine-tune one small model on four dead diarists. Give it a date; a ghost writes the entry. With a persona token you summon one writer; without it, a blended composite of all four speaks.

The ghosts: Samuel Pepys (1660s London) · Vincent van Gogh (1880s letters) · Katherine Mansfield (1910s–20s journal) · Mary MacLane (1902/1917 confessions). All US public domain — see PROVENANCE.md.

Quick start (RunPod or any CUDA box)

git clone <this repo> && cd Ghostwriter
pip install -r requirements.txt

# data is already built (data/dataset/), but to rebuild from raw:
python scripts/clean_corpora.py
python scripts/build_dataset.py

# train — Qwen2.5-3B full fine-tune, ~minutes on an A100:
python scripts/train.py --config configs/train_qwen3b.yaml
# (smaller GPU? configs/train_qwen3b_lora.yaml)

# hold the séance:
python scripts/sample.py --model outputs/ghost-qwen3b --date "7 June 2026"
python scripts/sample.py --model outputs/ghost-qwen3b --persona maclane \
    --date "7 June 2026" --seed-text "Today was hard."
python scripts/sample.py --model outputs/ghost-qwen3b \
    --prompt "What do you keep refusing to look at?"

No GPU handy? python scripts/train.py --config configs/smoke_test.yaml runs the entire pipeline on CPU with a 135M model in a few minutes.

Docs

• SPEC.md — what this is, how it works, next steps
• HOW_TO.md — raise your own ghost (general recipe + the copyright séance rules)
• CLAUDE_IDEAS.md — implemented extras and the speculative backlog
• PROVENANCE.md — sources, editions, and the ghosts we couldn't legally summon

Every training run prints its full hyperparameters, and post-train test inference prints prompt, output, and all generation settings — runs are self-documenting by design.