# Farka measured examples

Show the base model. Train. Test the tuned model on unseen data.

Farka treats artifact creation as an intermediate state. Promotion requires
paired base-versus-tuned evidence on a separate holdout, valid tuned output,
reconciled provider usage, and safe compute teardown.

## Qwen2.5 Coder 7B, code specialization

A light tune on one 16 GB accelerator. One pass over 270 training examples,
then tested on 30 examples kept out of training.

| Measure | Value |
| --- | --- |
| Base | 15/30 |
| Tuned | 30/30 |
| Change | +15 recovered on 30 unseen tasks |

SDK export, tuned inference, fleet promotion, settlement, and verified compute
teardown all passed for this run.

## SmolLM2 360M, small-model full tune

600 training examples, 60 for choosing the best checkpoint, and 60 blind test
examples the model never saw during training.

| Measure | Value |
| --- | --- |
| Base | 21/60 |
| Tuned | 58/60 |
| Output validity | 100% |

The best checkpoint was selected from paired development evidence. The
deployment bundle was produced only after promotion passed. Earlier failed
recipes are retained as negative evidence rather than discarded.

## Granite 4.1 8B, upper validated boundary

A 4-bit QLoRA run completed three epochs and all 102 optimizer steps on one
16 GB accelerator without an out-of-memory error.

| Measure | Value |
| --- | --- |
| Base | 0/30 |
| Tuned | 30/30 |
| Change | +30 recovered on the identical hidden holdout |

Weights export, tuned invocation, settlement, fleet promotion, and verified
compute teardown passed.

## DeepSeek R1 Distill Qwen 7B, reasoning-model specialization

The 7.6B dense model completed three 4-bit QLoRA epochs and 102 optimizer steps
on one 16 GB accelerator.

| Measure | Value |
| --- | --- |
| Base | 0/30 valid and correct |
| Tuned | 30/30 valid and correct |
| Change | +30 on the identical hidden holdout |

The deployment bundle, tuned invocation, settlement, fleet promotion, and
verified compute teardown passed.

## Phi-4 Mini 3.8B, compact reasoning model

A three-epoch 4-bit QLoRA run completed all 102 optimizer steps on one 16 GB
accelerator.

| Measure | Value |
| --- | --- |
| Base | 20/30 |
| Tuned | 30/30 |
| Change | +10 on the identical hidden holdout |

Weights export, tuned invocation, settlement, fleet promotion, and verified
compute teardown passed.

These are measured single-accelerator runs. 80 GB and multi-GPU training remain
unpublished until a live metered proof passes.

## Method

1. Profile and validate the dataset.
2. Quote before any compute starts.
3. Reserve credits and train.
4. Compare base and tuned models on the same unseen holdout.
5. Unlock delivery only if promotion passes.

## Contact

edi.hasaj@farka.ai