LoRA fine-tuning · cost intelligence
Stop overpaying for
LoRA fine-tunes.
Submit your dataset and base model. TuneBench sweeps a matrix of LoRA configs, measures cost, wall-clock, and eval quality — and tells you the cheapest config that hits your quality bar.
Free tier at launch · no card required · or try the free planner now →
| # | config | eval | cost | $/quality-pt |
|---|---|---|---|---|
| 1 | r=16 α=32 lr=2e-4 3epRECOMMENDED | 87.4 | $1.12 | $0.0128 |
| 2 | r=32 α=64 lr=2e-4 3ep | 88.1 | $1.87 | $0.0212 |
| 3 | r=8 α=16 lr=3e-4 4ep | 84.9 | $0.94 | $0.0221 |
| 4 | r=64 α=128 lr=1e-4 3ep | 88.6 | $3.41 | $0.0385 |
| 5 | r=16 α=32 lr=1e-4 5ep | 86.2 | $2.29 | $0.0412 |
analyst: r=16 hits your 85 bar at 33% of the cost of r=64. The extra 1.2 eval points from r=64 cost you 3× — not worth it for this dataset.
Run the whole matrix
One submission fans out to a grid of rank, alpha, learning-rate, and epoch combinations. No more one-config-at-a-time guesswork in a notebook.
Cost-per-quality-point ranking
Every config reports wall-clock, dollar cost, and eval score — then gets ranked by what a quality point actually costs you. Accuracy alone is a vanity metric.
A recommendation, not a spreadsheet
Set your quality bar and an LLM analyst reads the full matrix and tells you the cheapest config that clears it — with the reasoning written out.
How it works
Submit
Point us at your dataset and base model. Set the quality bar you need to hit.
Benchmark
We sweep the LoRA config matrix and record cost, wall-clock, and eval quality per cell.
Ship the cheapest winner
Get the config that hits your bar at the lowest cost — copy the exact training command.
Pricing
Compute passthrough included in run quotas. Prices in USD and INR.
Pro
POPULAR$19₹1,599 / mo
For solo builders
- 20 runs / month
- Private leaderboard
- LLM config recommendations
- Email reports
Team
$49₹3,999 / mo
For small ML teams
- Everything in Pro
- Shared dashboards
- 3 seats included
- Priority queue
FAQ
- What does TuneBench actually run?
- A matrix of LoRA fine-tuning configurations (rank, alpha, learning rate, epochs) against your dataset and base model, each measured for wall-clock time, compute cost, and eval quality.
- What is cost-per-quality-point?
- Total run cost divided by eval score. It surfaces configs that are nearly as good as the top scorer at a fraction of the price — usually a 40–70% saving over defaulting to the biggest rank.
- Which base models are supported?
- Open-weight models at launch: Llama, Mistral, Qwen, Gemma families. Anything you can LoRA-tune with PEFT.
- Do I need my own GPUs?
- No. Runs execute on managed cloud GPUs and the cost you see is the cost you pay. You can also import results from your own runs to rank them.