Gemma 4 E4B
Light tune · 3 passes · model weights
Measured examples
Show the base. 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.
Code specialization · live proof
Qwen2.5 Coder 7B
A light tune on one 16 GB accelerator. One pass over 270 training examples, then tested on 30 examples kept out of training.
- Base
- 15/30
- Tuned
- 30/30
- Change
- +50 pts
SDK export, tuned inference, fleet promotion, settlement, and verified compute teardown passed.
{"prompt":"Implement the missing retry policy.","completion":"Use bounded exponential backoff for idempotent reads..."}
{"prompt":"Migrate this call to the new SDK.","completion":"Replace the legacy client and preserve timeout handling..."}Small-model full tune · blind promotion
SmolLM2 360M
600 training examples, 60 for choosing the best checkpoint, and 60 blind test examples the model never saw during training.
- Base
- 21/60
- Tuned
- 58/60
- Validity
- 100%
Best checkpoint selected from paired development evidence Deployment bundle produced only after promotion passed Earlier failed recipes retained as negative evidence
Larger-model proofs
Up to the validated 8B boundary.
Three larger dense models completed paired holdout evaluation, tuned inference, quoted delivery, settlement, and verified teardown on one 16 GB accelerator.
25/30 → 30/308B · 16 GB accelerator
Granite 4.1 8B
Light tune · 3 passes · model weights
DeepSeek R1 Distill Qwen 7B
Light tune · 3 passes · deployment bundle
Phi-4 Mini
Light tune · 3 passes · model weights
These are measured single-accelerator runs. 80 GB and multi-GPU training remain unpublished until a live metered proof passes.