fine-tune-svc — launch kit

1-liner

LoRA fine-tuning service for SMBs. Customer provides CSV; we deliver a Qwen LoRA they deploy on their box. $2-10K per project.

Tweet hook

Fine-tuning got cheap in 2026. LoRA on Qwen 30B = 4-6 hours on a 3090. But getting from "I have data" to "I have a working LoRA" is still 2 weeks for most teams.

Built the pipeline. $2-10K/project.

Workflow 🧵

Cold-email ICP

• ML teams at startups + SMBs needing custom-tuned models
• B2B SaaS with their own user-interaction data wanting domain-specific bots
• Privacy-sensitive teams (legal, medical, finance) wanting on-prem fine-tunes

Cold-email template

Subject: $2K LoRA fine-tune for {their domain}

Hi {first} — for {company} ML/AI work.

Most teams that should fine-tune don't because the pipeline is painful.
We do it: data-prep + LoRA training + eval + deployable adapter.
$2-10K depending on dataset size + complexity.

Free 30-min consult. Reply with the task you'd want fine-tuned for.

SEO content

1. "LoRA economics 2026: when fine-tuning beats prompting"
2. "Unsloth + Qwen LoRA stack — full setup"
3. "Custom-fine-tune ROI vs cloud prompt-engineering"

fine-tune-svc

LoRA fine-tuning service for SMBs. Customer provides a prompt+completion CSV; we deliver a fine-tuned Qwen LoRA adapter they deploy on their hardware (or our hosted endpoint).

Pricing

• $2,000-10,000 per project depending on dataset size + complexity
• $499/mo retainer for ongoing improvement passes
• $25K for full custom training (their own base, multi-LoRA, eval suite)

When fine-tuning beats prompting

• Heavy domain vocabulary (legal, medical specialty, internal jargon)
• 10K+ examples of the exact task
• Need consistent output structure that prompts can't reliably enforce
• Need lower latency / smaller model than baseline Qwen 30B

Run (dataset prep)

cd C:\openclaw-products\fine-tune-svc
python -m venv .venv
.\.venv\Scripts\activate
pip install -e .

# Prepare a CSV (columns: prompt, completion, system optional)
finetune prep customer-data.csv --out prepared/

The actual training step uses Unsloth or Axolotl on a GPU box (the operator's H100 / H200 / 3090). v0 is dataset prep only; the training runner is a script wrapper around Unsloth's CLI.

Roadmap

• Unsloth wrapper (full pipeline: data → fine-tune → eval)
• Eval-set scoring against the original Qwen base
• Hyperparameter search (LoRA rank, alpha, dropout)
• Hosted-endpoint deployment (the customer doesn't need their own GPU)
• Continual-learning mode (monthly refresh on new examples)