Fine-tuning project scoper

Answering from your documents

• Have you tried retrieval first? Most document questions are a retrieval problem, not a fine-tuning one.
• How often do the documents change, and must answers cite the passage they came from?
• If retrieval works but answers are badly shaped, would a small tuned model that reads retrieved passages be enough?

Format, tone or style

• Can you write 20 examples of the ideal output by hand? If not, the target is not yet clear enough to train on.
• Does a strong base model with a few examples in the prompt already get close? Measure before training.
• Who signs off that an output is right, and can that judgment be written down as a rubric?

Classification, routing or extraction

• How many labels or fields, and are they stable? A changing label set means retraining.
• What error rate is acceptable, and which mistakes cost the most?
• At your daily volume, would a small tuned model be cheaper than a large model with a long prompt?

Multi-step tasks with tools

• Can success be checked automatically (tests pass, the right record changed)? That makes reinforcement fine-tuning possible.
• Do you have recorded traces of good runs to learn from, or only end results?
• Where will the agent run during training: a copy of your systems or a simulated environment?

Preferring some answers over others

• Can reviewers rank two answers to the same input consistently? Preference tuning (DPO) needs ranked pairs.
• Will you tune on top of a supervised model first? Preference tuning usually follows SFT.
• Who reviews the pairs, and how will you check reviewers agree?

With no labeled examples yet

• Start with prompting and an evaluation set of real inputs. Training comes later.
• Who will write or label the first examples: your staff, a data vendor or both?
• Can any examples be generated from existing logs without personal details?

With fewer than 300 examples

• Set aside at least a fifth as a held-out test set before training anything.
• Is a parameter-efficient method (LoRA) enough? It is cheaper and less prone to overfitting small sets.
• Check the job will clear the provider's minimum charge without padding the data.

With roughly 300 to 3,000 examples

• How many of them have been checked by a person? A smaller clean set beats a larger noisy one.
• Do they cover the rare cases, or only the common ones?
• How many epochs will you run? Training tokens are billed per epoch.

With more than 3,000 examples

• Is full fine-tuning worth it over LoRA for your gain in quality? Some platforms price full tuning at twice LoRA.
• Do you need all of them, or would a filtered subset train as well for less?
• Is distilling from a larger model an option to reach a smaller, cheaper one?

On the provider's shared serving

• Does the provider serve tuned models on shared capacity, or only on dedicated hardware?
• Are tuned-model token prices the same as the base model, or higher in your region?

On dedicated GPUs

• What is the billing unit (per second, minute or hour) and does it run when idle?
• Can the deployment scale to zero overnight?
• How many GPUs does your model size need at your peak traffic?

In your own cloud account

• Can the tuned weights be exported, and in what format?
• Which inference stack will you run, and who maintains it?
• Does your security review need training as well as serving inside your account?

On your own hardware

• Do you have GPUs with enough memory for the model size you picked?
• Is the base model license fine for on-premise commercial use?
• Who updates and monitors the model after launch?

Small open model (up to 16B)

• Fireworks AI: $0.50 (LoRA SFT) to $2.00 (full DPO) per 1M training tokens for models up to 16B.
• Together AI: $0.34 SFT and $0.84 DPO per 1M tokens for Llama 3.1 8B, $4.00 minimum per job.
• Google Vertex AI: from $0.28 (Llama 3.2 1B) to $8.46 (Qwen 3 14B) per 1M training tokens.

Mid-size open model (16B to 80B)

• Fireworks AI: $3.00 (LoRA SFT) to $12.00 (full DPO) per 1M training tokens for 16.1B to 80B.
• Together AI: $2.03 SFT and $5.08 DPO per 1M tokens for Llama 3.3 70B.
• Google Vertex AI: $6.72 per 1M training tokens for Llama 3.3 70B; $6.57 for Qwen 3 32B.

Large open model (80B to 300B)

• Fireworks AI: $6.00 (LoRA SFT) to $24.00 (full DPO) per 1M training tokens for 80B to 300B.
• Together AI: $2.50 SFT per 1M tokens for gpt-oss-120B; $4.20 for Qwen3.5 122B A10B.

Very large open model (over 300B)

• Fireworks AI: $10.00 (LoRA SFT) to $40.00 (full DPO) per 1M training tokens for models over 300B.
• Together AI: $5.60 SFT and $14.00 DPO per 1M tokens for Qwen3.5 397B A17B, $22.00 minimum per job.

A closed cloud model (Gemini, GPT)

• Google Vertex AI: Gemini tuning from $0.0015 per 1,000 training tokens (Gemini 2.5 Flash Lite) to $0.025 (Gemini 2.5 Pro).
• Azure OpenAI: lists GPT-4.1 family and GPT-4o tuning; figures not captured.
• OpenAI: retiring fine-tuning, no new customers (its price list, 7 Nov 2026).

Provider's shared serving

Vertex AI bills tuned Gemini 3+ models at the base model price in the global region (1.5x in other regions).

Dedicated GPUs from the provider

Fireworks AI bills deployments per GPU-second with no extra charge for a tuned model; Together AI bills dedicated endpoints per minute by hardware.

Our own cloud account

Your cloud bill for GPUs and storage applies; check the provider allows exporting tuned weights.

Our own hardware

Hardware and staff costs are yours; the training run itself can still be bought from a platform if weights can be exported.