Fine-tune an LLM to match a writing style with Together AI

The tg command uploads data, starts jobs and deploys models, so every later step runs from one terminal.

1. 1Run the command below to install the CLI with uv.
2. 2Run tg login and approve the sign-in in your browser.
3. 3Copy a key from the API keys page of your Together account, then run export TOGETHER_API_KEY=[YOUR KEY] for the test in step 6.

uv tool install "together[cli]" --upgrade

You should see the sign-in approved in your browser, and no error in the terminal.

Keep your API key out of scripts, notebooks and shared files. Set it in the terminal, as above, instead.

The model learns a style from pairs: a short request, then a passage from your text as the answer.

1. 1Put your text in a folder named corpus, as .txt files.
2. 2Delete what isn't prose, such as page numbers, tables of contents, indexes and running headers.
3. 3Save the script below as make_data.py next to the folder, change [STYLE NAME], and run python3 make_data.py.

import json, pathlib

STYLE = "[STYLE NAME]"
examples = []
for f in sorted(pathlib.Path("corpus").glob("*.txt")):
    paras = [p.strip() for p in f.read_text(encoding="utf-8").split("\n\n") if len(p.split()) > 20]
    chunk = []
    for p in paras:
        chunk.append(p)
        if sum(len(c.split()) for c in chunk) >= 300:
            text = "\n\n".join(chunk)
            topic = text.split(". ")[0][:150]
            examples.append({"messages": [
                {"role": "user", "content": f"Write a passage in the style of {STYLE} about: {topic}"},
                {"role": "assistant", "content": text}]})
            chunk = []

with open("train.jsonl", "w", encoding="utf-8") as out:
    for e in examples:
        out.write(json.dumps(e, ensure_ascii=False) + "\n")
print(len(examples), "examples written to train.jsonl")

You should see a count of examples, and a train.jsonl file with one example per line.

Together warns that training a reasoning model on examples without reasoning steps can weaken its reasoning. For a writing model, that matters less.

The check runs on your computer, so format errors show up before you pay for anything.

1. 1Run the check below.
2. 2Upload the file with tg files upload train.jsonl.
3. 3Copy the file ID from the upload output.

tg files check train.jsonl

You should see a check report with no errors, then a file ID after the upload.

LoRA trains a small add-on for the base model, which costs far less than retraining the whole model.

1. 1Put your file ID into the command below and run it.
2. 2Read the price estimate the CLI prints, then confirm to start the job.
3. 3Copy the job ID, and follow progress with tg fine-tuning list-events [JOB_ID].

tg fine-tuning create --training-file [FILE_ID] --model Qwen/Qwen3.5-9B --lora --lora-r 16 --lora-alpha 32 --n-epochs 1 --suffix my-style

You should see a price estimate, then a job ID once you confirm.

You can close the terminal while the job runs. tg fine-tuning retrieve [JOB_ID] shows its status at any time.

Tuned models run on a dedicated endpoint, not on Together's per-token API.

1. 1When the job has finished, run tg fine-tuning retrieve [JOB_ID] and copy the model object name.
2. 2Run the command below with that name.
3. 3Run tg endpoints get [ENDPOINT_ID] until the state reads DEPLOYMENT_STATE_READY.
4. 4Copy the endpoint string the deploy command printed, such as your-project/my-style.

tg endpoints deploy "[MODEL OBJECT NAME]" --endpoint my-style

You should see the state DEPLOYMENT_STATE_READY, and an endpoint string to call.

The endpoint bills by the minute while it runs, even when idle.

The same prompts on both models show what the training changed.

1. 1Send 3 prompts with the command below: a topic your text covers, a made-up topic, and a topic newer than your text.
2. 2Run the same 3 prompts on the base model with ollama run qwen3.5:9b.
3. 3When you're done, delete the endpoint with tg endpoints rm [ENDPOINT_ID] --force.

curl https://api-inference.together.ai/v1/chat/completions -H "Authorization: Bearer $TOGETHER_API_KEY" -H "Content-Type: application/json" -d '{"model": "[YOUR ENDPOINT STRING]", "messages": [{"role": "user", "content": "Write a passage in the style of [STYLE NAME] about: [YOUR TOPIC]"}], "max_tokens": 600}'

You should see the tuned model writing like your text, while the base model writes in its usual voice.

Use the same 3 prompts after every new run, so you can tell whether a change helped.