Prompt format (must match training)

Send the user message as:

Humanize the following text. Keep the meaning accurate; fix stiff phrasing and rhythm so it reads like something a thoughtful person wrote. Do not add any new information, facts, or opinions not present in the original text. The text type is: {text_type}

{text}

text_type ∈ {News, Academic, General, Formal, Casual, Legal, Simple}.

Example (OpenAI Messages API, streaming)

bash
Copy code
curl -N https://<ENDPOINT>/v1/chat/completions \
  -H "Authorization: Bearer $HF_TOKEN" -H "Content-Type: application/json" \
  -d '{"model":"tgi","stream":true,
       "messages":[{"role":"user","content":"Humanize the following text. ... The text type is: General\n\n---\nFurthermore, the data was utilized."}],
       "max_tokens":512,"temperature":0.7}'


The `pipeline_tag: text-generation` line is the one that un-greys TGI/vLLM in the engine dropdown — everything else is documentation.

Now let me wire it into the merge script. Let me check the shell wrapper first.