Model Fine Tuning
A model is only as good as its data.
The tuning practice starts long before training: build the corpus, prove the corpus, then touch the weights. This page is the process — corpus construction and qualification at six-figure scale, labelling calibrated before it is trusted, and fine-tuning runs executed on local hardware, start to finish.
Corpus building & qualification
The flagship programme is the corpus behind a 4B vulnerability-triage model: assembled in-house, decontaminated against held-out evaluation data, and passed through a qualification gate before a single training run was allowed to consume it. The gate earned its keep — it caught 604 leaked held-out rows inside the training file, rows that would have silently inflated every downstream measurement.
| Training corpus | 140,545 rowsAssembled and decontaminated in-house |
|---|---|
| Rebuilt corpus | 108,978 rowsThe current frozen candidate lineage |
| Leaked held-out rows caught | 604Found by the decontamination check before training |
| Labelling | Model-adjudicated, calibratedThousands of rows labelled by model judgement, with calibration draws scored before any gate verdict counts |
| Evaluation discipline | Frozen public benchmarks onlyEvery run measured against the same frozen runner — no moving targets |
The runs
| LoRA fine-tune — 4B base | 37.03 hoursLoRA r=32, α=64 on Qwen3-4B-Instruct-2507 · 140,545 rows · 1 epoch · 79,830,205 tokens · trained on a DGX Spark (GB10) |
|---|---|
| Step-distillation LoRAs — video | 20 steps → 4Turbo adapters for MiniMax H3: FL2V 4-step, FL2V 8-step, Ref2V 4-step (~1.96 GB each), deployed against the 20-step baseline |
| FLUX.1-schnell LoRA | 500 steps, completeai-toolkit run with final and step-250 adapters kept on disk |
Process and scale numbers only, deliberately: what gets published for work in progress is how it was built and how it was checked — the corpus receipts, the gate verdicts, the run configs — not interim scores.