# Fix a voice, persona or output format that prompting doesn't hold

> Fix a voice, persona or output format that prompting doesn't hold: 2 companies fully do it, 7 with a limit, 13 carry the tag. Every listing in Fine-tuning with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/fine-tuning.persona-style/
- Markdown: https://www.anchorterminal.com/niches/fine-tuning.persona-style/index.md (~1,050 tokens)
- Slim: https://www.anchorterminal.com/niches/fine-tuning.persona-style/index.min.md (~330 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/fine-tuning.persona-style/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Two or three companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 2 | 7 | 13 | 5 | 3 |

## What counts as done

Supervised tuning on the examples is documented for the chosen base model, and the docs give a small-dataset recipe (example count, epochs or learning rate) for style or persona work.

How agents ask for it: make the model always answer in this character's voice; train it to write like me; make it follow our format without a long prompt

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Unsloth (Unsloth)**: SFT with LoRA is the core path, and the hyperparameters guide recommends 1 to 3 epochs, explains learning rate choices and gives rank and alpha settings for small runs.
- **Vertex AI Gemini tuning (Google LLC)**: The tuning overview recommends SFT for strict formatting with hundreds to thousands of examples, and epoch count and adapter size are documented, with epochs defaulting to a value fitted to the dataset.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Amazon Bedrock model customisation (Amazon Web Services, Inc.)**: SFT is documented per model with a hyperparameters page, but the only example-count guidance read (start with 100 to 200 examples) is for RFT, not style tuning.
- **Axolotl (Axolotl AI)**: SFT on conversation datasets with assistant-only loss is documented with per-model example configs, but no small-dataset recipe for style or persona was found.
- **Microsoft Foundry fine-tuning (Azure OpenAI) (Microsoft Corporation)**: SFT with service-selected hyperparameter defaults is documented and the cost example uses 500 prompt-response pairs, but no style or persona recipe with epochs or learning rate was found.
- **Fireworks AI Fine-tuning (Fireworks.ai, Inc.)**: SFT runs through the cookbook sft_loop and a free Muse Glimmer 30B first fine-tune, but no small-dataset recipe for style or persona was found, and the agent must keep the training process running.
- **Nebius Token Factory fine-tuning (Nebius B.V.)**: The SFT guide gives example hyperparameters (n_epochs 3, lora_r 16) and a 10 to 20 per cent validation split, but no example count or recipe for style or persona work.
- **Tinker (Thinking Machines Labs, Inc.)**: A Chat SFT recipe trains on conversational datasets, but no small-dataset recipe for style or persona appears in the README; acc-42 made a model with SFT only, and acc-43 moved to RAG on cost.
- **Together AI Fine-tuning (Together Computer, Inc.)**: SFT is the default method with documented data formats and early stopping, but no small-dataset recipe (example count, epochs, learning rate) for style or persona work was found in the pages read.

## What stops agents finishing

- training cost and serving cost above what a prompt and RAG achieve

## Around it

After: inference.tool-loop, personal-assistants.chat.

## Where this comes from

Weight 5 of 100 in Fine-tuning: 3 of 47 accounts (a Tinker persona model, a character-creation user on the OpenAI forum, and a personal clone that went back to RAG); low count, kept separate because the success test differs

As data: https://www.anchorterminal.com/api/v1/use-cases/fine-tuning.persona-style.json, or the MCP tool get_use_case with id fine-tuning.persona-style. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
