Humanizer is a local-first rewrite workspace for long English text. It splits text into semantic chunks, generates left_skew, middle_skew, and right_skew alternatives, and lets you assemble a cleaner final draft chunk by chunk.
This public repository now exposes one unique runnable baseline so users do not have to guess which model stack was actually used.
- Base model:
TinyLlama/TinyLlama-1.1B-Chat-v1.0 - Adapter version:
humanizer-tinyllama-v6-step-0040 - Adapter download location: GitHub Release asset on
kevin0818-lxd/Humanizer - Model server:
mistralrsOpenAI-compatible HTTP - Expected model port:
8081 - Expected app API port:
8000
git clone https://github.com/kevin0818-lxd/Humanizer.git
cd HumanizerDownload the Hugging Face base model into:
data/models/tinyllama-1.1b-chat-v1.0
Recommended source:
TinyLlama/TinyLlama-1.1B-Chat-v1.0
Example:
huggingface-cli download TinyLlama/TinyLlama-1.1B-Chat-v1.0 \
--local-dir data/models/tinyllama-1.1b-chat-v1.0Download the release asset:
humanizer-tinyllama-v6-step-0040.zip
Unzip it so the adapter directory becomes:
data/models/humanizer-tinyllama-v6-step-0040
After unzip, that folder should contain:
adapter_model.safetensorsadapter_config.jsonmetrics.jsonmanifest.jsonREADME.md
cp .env.example .env
uv sync --directory apps/api --extra dev
npm --prefix apps/web installmistralrs serve \
-m data/models/tinyllama-1.1b-chat-v1.0 \
--lora data/models/humanizer-tinyllama-v6-step-0040 \
--max-seq-len 4096 \
--max-batch-size 1 \
--host 0.0.0.0 \
--port 8081 \
--no-uiuv run --directory apps/api uvicorn app.main:app --host 127.0.0.1 --port 8000npm --prefix apps/web run dev- Model server:
http://127.0.0.1:8081 - App API:
http://127.0.0.1:8000 - Web UI:
http://127.0.0.1:3000
curl http://127.0.0.1:8000/healthExpected:
{"status":"ok","mode":"http"}curl http://127.0.0.1:8081/v1/modelsExpected:
- one loaded local model
- model id
default
Open:
http://127.0.0.1:3000
Paste a paragraph, generate candidates, and confirm that the app returns real left_skew, middle_skew, and right_skew outputs instead of mock text.
apps/api: FastAPI backendapps/web: Next.js frontendscripts: dataset, training, evaluation, and probing utilitiesdocs: dataset and training design notesconfigs: local experiment configs
This repository intentionally does not include:
.envfiles or private credentials- local databases and runtime logs
- uploaded source files and derived datasets
- downloaded models, adapters, or checkpoints in Git history
- local evaluation artifacts and caches
- machine-specific workspace files
The runnable adapter is published as a GitHub Release asset, not in the main repository history.