Data must be usable
Farka accepts structured formats such as JSONL, JSON, and CSV, but format alone does not prove that examples are consistent or that you have the right to train on them. The dataset is validated before the run.
Fine-tuning field guide
Fine-tune an LLM on your own data. Then prove it helped.
A specialized model starts with a narrow job, representative examples, and a test the model has not seen. Farka keeps those decisions attached to the quote, the training run, and the artifact you receive.
Start with a test
Define improvement before touching the weights.
“Fine-tune my LLM” is not yet a training plan. Name the behavior you want: classify a support request, follow a required output format, or answer from a bounded domain. Keep some examples out of training and decide how you will score them. Otherwise a model can repeat the training examples and appear to improve without doing the new job well.
If the task can be solved with better retrieval or a clearer prompt, training may not be the right first move. The evaluation makes that decision visible.
From dataset to delivery
Five decisions in the right order.
- 01
Name the job
Choose one task you can judge. Set aside examples the model will never see during training, so a later score means something.
- 02
Prepare your examples
Upload a structured dataset. Farka validates and profiles it before a paid run. Check your rights to the data and keep the held-out examples separate.
- 03
Choose and quote
Pick a supported, license-reviewed model and delivery target. The quote fixes the workload and shows the expected cost before you fund or launch compute.
- 04
Train and compare
Run the approved job, then ask the same unseen questions of the base and tuned models. A finished training job is not automatically a better model.
- 05
Take the artifact
A run that passes its quality checks can unlock model weights or a runnable deployment bundle for infrastructure you operate.
One measured example
Compare the base and tuned model on the same unseen work.
In Farka's published Qwen2.5 Coder 7B run, a light tune used 270 training examples and one 16 GB accelerator. The model then faced 30 examples kept out of training. It answered 15 correctly before the tune and 30 after it. That result belongs to this configuration; it is not a promise for another dataset or model.
Inspect the measured examples- Base model
- 15/30
- Tuned model
- 30/30
- Unseen test tasks
- 30
Cost precedes compute
A quote shows the model, data, compute, region, and delivery choice. You can inspect an estimate before funding credits. Launch reserves the approved cost and starts the training workflow.
Delivery has a boundary
Choose open weights or a runnable Docker bundle before paying. The bundle's inference routes run in your infrastructure. Farka-managed hosted inference is not offered in the current launch.
Common decisions
When should you use the custom path?
If your candidate model is over the validated self-service size, uses a sparse or mixture-of-experts architecture, needs several accelerators, or must run inside a private network, request a capacity and security review. Do not infer support or quality from a model's presence in a catalog. Farka reviews the complete configuration and keeps unverified work out of the self-service lane.
Bring a task and a dataset